Aerial VLN-oriented annular route guiding method and related equipment
By fusing satellite imagery with vector map features and optimizing path segments, multi-granularity route guidance instructions are generated, solving the problems of path deviation and poor stability in existing airborne VLN navigation and achieving accurate and stable airborne VLN navigation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HONG KONG UNIV OF SCI & TECH (GUANGZHOU)
- Filing Date
- 2026-03-10
- Publication Date
- 2026-05-19
AI Technical Summary
Existing airborne VLN route guidance methods fail to effectively combine visual semantic features and global geometric maps, resulting in navigation path deviations and poor stability. Furthermore, they do not take into account human visual cognitive habits and UAV flight constraints, making it difficult to meet the navigation needs of complex aerial scenarios.
By fusing satellite imagery with vector map features, a cognitive saliency map and path comprehensibility score are generated. The path is segmented in conjunction with UAV dynamic constraints, and multi-granularity route guidance instructions are generated, thus achieving the unification of visual perception and flight constraints.
It improves the navigation accuracy, stability, and environmental adaptability of the aerial VLN system, ensuring that the path is visually clear and easy to understand, conforms to human cognitive logic, and is adapted to the performance of the UAV.
Smart Images
Figure CN122062691A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of route planning technology, and more specifically, to a circular route guidance method and related equipment for airborne VLN. Background Technology
[0002] With the deep integration of UAV technology and Visual Language Navigation (VLN) technology, aerial VLN is increasingly widely used in disaster relief, geographic surveying, logistics and transportation. Circular routes have become one of the core application scenarios of aerial VLN due to their advantages such as full coverage observation, ease of return and review, and suitability for long-term aerial observation.
[0003] Current air navigation methods for VLN (Vehicle-to-Noise Navigation) still have many shortcomings and cannot meet the needs of practical applications. The mainstream air navigation technologies fall into two categories: one focuses on geometric path planning, generating paths solely based on geometric information such as terrain and boundaries in vector maps, without fully considering the visual semantic features of the aerial environment. This results in paths that, while geometrically plausible, are disconnected from the actual aerial visual scene, making it difficult for UAVs to match paths visually during navigation and leading to deviations from the flight path. The other category relies on real-time ground imagery collected by the UAV for visual navigation, without incorporating the constraints of a global geometric map. This makes it highly susceptible to factors such as changes in ambient lighting, obstacle obstruction, and weather interference, resulting in poor navigation stability. Furthermore, in complex airspace environments, it is prone to chaotic path planning and the inability to form standardized circular routes. Furthermore, existing circular route guidance methods only focus on the physical feasibility of the route during the route selection process, without considering human visual perception habits. The generated routes lack visual recognizability, which is not conducive to the matching and execution of subsequent visual language commands. At the same time, the design of route control and guidance commands is relatively crude and has not been refined by combining the flight constraints of the UAV itself, resulting in insufficient navigation accuracy. This makes it difficult to adapt to complex aerial navigation scenarios with long field of view and high dynamics, and fails to meet the high efficiency and accuracy requirements of aerial VLN for circular route guidance.
[0004] Therefore, there is an urgent need for a new circular route guidance method for airborne VLNs to overcome the shortcomings of existing technologies, achieve accurate and stable guidance for airborne VLN circular routes, and meet the practical application needs of complex aerial scenarios. Summary of the Invention
[0005] This application provides a circular route guidance method and related equipment for airborne VLNs. By feature fusion of satellite imagery and vector maps, cognitively guided target circular path selection, constraint-guided path segmentation, and multi-granularity guidance instruction generation, it achieves accurate guidance for airborne VLN circular routes, improving the stability, reliability, and environmental adaptability of the navigation system.
[0006] A circular route guidance method for airborne VLNs includes:
[0007] Acquire satellite imagery and vector map of the target area. Using the geometric structure of the vector map as a constraint, map and align the visual features of the satellite imagery to the geometric feature space of the vector map to generate a fused feature representation.
[0008] Closed boundary elements are extracted from the vector map as candidate circular paths. Based on the fusion feature representation, the saliency distribution and structural comprehensibility of each candidate circular path at the visual perception level are analyzed to generate a cognitive saliency map and a path comprehensibility score. Target circular paths that meet the preset cognitive preference threshold are then selected based on the path comprehensibility score.
[0009] By combining the dynamic constraints of the UAV with the cognitive saliency map, the cognitive decision points and physical constraint points on the target circular path are identified, and the target circular path is divided into multiple continuous small segments using the cognitive decision points and the physical constraint points as segmentation boundaries.
[0010] Based on the local fusion feature representations of each of the aforementioned road segments and their corresponding local fusion feature representations, as well as the global fusion feature representations of the entire target loop path, a multi-granularity route guidance instruction containing local action guidance and global semantic description is generated.
[0011] Optionally, the method of mapping and aligning the visual features of the satellite imagery to the geometric feature space of the vector map includes:
[0012] In data-sparse regions, supplementary image data consistent with the distribution of real images is synthesized based on generative adversarial methods.
[0013] By using a contrastive learning approach, the geometric topology of the vector map is used as the positive sample anchor point to maximize the correlation between image blocks and vector structures at the same location in the feature space, thereby achieving semantic mapping from visual features to geometric feature space.
[0014] Optionally, the process of filtering to obtain the target circular path further includes:
[0015] Visual and morphological feature indicators of candidate circular paths are extracted, and the target circular path is obtained by comprehensive screening based on the visual feature indicators, the morphological feature indicators, the cognitive saliency map and the path comprehensibility score.
[0016] The visual feature index includes at least one of the following: internal homogeneity index, boundary sharpness index, and landmark availability index. The internal homogeneity index is used to measure the consistency of texture and color within the area surrounded by the path. The boundary sharpness index is used to evaluate the clarity of the path boundary in the image. The landmark availability index is used to identify significant landmark elements present around the path.
[0017] The morphological feature index includes at least one of the path length index, shape index index, and length ratio index, wherein the path length index is used to measure the overall length of the path, the shape index index is used to characterize the geometric complexity and morphological features of the path, and the angle-to-length ratio index is used to characterize the relationship between the path turning angle and the segment length to determine the executability of the action command.
[0018] Optional, also includes:
[0019] The visual feature indicators, morphological feature indicators, cognitive saliency map, path comprehensibility score, and UAV dynamic constraints are weighted and fused to generate a flightability score for each candidate circular path;
[0020] The candidate loop paths are sorted according to the flightability score, and the candidate loop paths whose scores meet the preset threshold are selected as the target loop paths.
[0021] Optionally, by combining the UAV dynamic constraints with the cognitive saliency map, cognitive decision points and physical constraint points on the target circular path are identified, and the target circular path is divided into multiple continuous small segments using the cognitive decision points and the physical constraint points as segmentation boundaries, including:
[0022] The target circular path is geometrically simplified, and key corner points on the path are extracted;
[0023] The target circular path is initially segmented using the key corner points to obtain an initial set of road segments;
[0024] Peak detection is performed on the cognitive saliency map to identify the path locations corresponding to the local maxima of attention weights, which are then used as cognitive decision points.
[0025] Based on the preset minimum turning radius constraint of the UAV, the curvature radius of each point on the path is calculated, and the positions with curvature radii less than the minimum turning radius are marked as physical constraint points;
[0026] Using the cognitive decision points and the physical constraint points as optimization boundaries, the initial road segment set is divided into two parts. Points where the cognitive decision points and physical constraint points overlap or are less than a preset threshold will be used as mandatory segmentation points for the second division.
[0027] For road segments whose length exceeds a preset threshold after secondary division, they are uniformly divided at a fixed step size to form the final set of road segments.
[0028] Optionally, the geometric simplification of the target circular path and the extraction of key corner points on the path includes:
[0029] The Douglas-Peucker algorithm is used to geometrically simplify the target circular path, and the vertices of the simplified polyline are extracted as key corner points.
[0030] Optionally, based on each of the aforementioned road segments and their corresponding local fusion feature representations, as well as the global fusion feature representation of the entire target loop path, a multi-granularity route guidance instruction containing local action guidance and global semantic description is generated, including:
[0031] For each road segment, based on the road segment and its corresponding local fusion feature representation, the boundary geometric features and surrounding land feature distribution of the road segment are analyzed to generate local route guidance describing the navigation mode of the road segment;
[0032] Based on the global fusion feature representation of the entire target ring path and surrounding area, the semantic structure and spatial relationships of the overall scene are analyzed to generate a global semantic description of the overall navigation mission.
[0033] The local route guidance and the global semantic description are semantically integrated and structured to generate multi-granularity route guidance instructions.
[0034] Optionally, after generating the multi-granularity route guidance instructions, the method further includes:
[0035] A UAV dynamics simulation environment is constructed, and the generated multi-granularity flight path guidance instructions are parsed into a sequence of control instructions and executed in the simulation environment. The spatiotemporal deviation between the actual flight trajectory and the planned path is collected.
[0036] Based on the spatiotemporal deviation, the instruction generation strategy is iteratively optimized until the deviation converges to a preset threshold, and the final executable route guidance instruction is output.
[0037] A circular route guidance device for airborne VLNs, comprising:
[0038] The data fusion module is used to acquire satellite imagery and vector maps of the target area, and using the geometric structure of the vector map as a constraint, to map and align the visual features of the satellite imagery to the geometric feature space of the vector map to generate a fused feature representation.
[0039] The path filtering module is used to extract closed boundary elements from the vector map as candidate circular paths, analyze the saliency distribution and structural comprehensibility of each candidate circular path at the visual perception level based on the fused feature representation, generate a cognitive saliency map and a path comprehensibility score, and filter out target circular paths that meet the preset cognitive preference threshold based on the path comprehensibility score.
[0040] The path segmentation module is used to combine the UAV dynamic constraints with the cognitive saliency map to identify the cognitive decision points and physical constraint points on the target circular path, and to divide the target circular path into multiple continuous small segments using the cognitive decision points and physical constraint points as segmentation boundaries.
[0041] The instruction generation module is used to generate multi-granularity route guidance instructions that include local action guidance and global semantic description based on each of the small road segments and their corresponding local fusion feature representations, as well as the global fusion feature representation of the entire target loop path.
[0042] A ring-shaped wayfinding device for airborne VLNs, comprising a memory and a processor;
[0043] The memory is used to store programs;
[0044] The processor is configured to execute the program to implement the various steps of the circular route guidance method for airborne VLNs as described in any of the preceding claims.
[0045] A readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the circular route guidance method for airborne VLNs as described in any of the preceding claims.
[0046] A computer program product includes a computer program that, when executed by a processor, performs the steps of the circular route guidance method for airborne VLNs as described in any of the preceding claims.
[0047] As can be seen from the above technical solutions, the circular route guidance method and related equipment for airborne VLN provided in this application acquires satellite imagery and vector maps of the target area. Using the geometric structure of the vector map as a constraint, the visual features of the satellite imagery are mapped and aligned to the geometric feature space of the vector map to generate a fused feature representation. This organically unifies visual semantic information and geometric spatial information, constructing an environmental representation that combines geometric accuracy and visual richness, providing more comprehensive and reliable data support for subsequent path analysis and route guidance. By extracting closed boundary elements from the vector map as candidate circular paths, and analyzing the saliency distribution and structural comprehensibility of each candidate circular path at the visual perception level based on the fused feature representation, a cognitive saliency map and a path comprehensibility score are generated. Based on the path comprehensibility score, target circular paths that meet the preset cognitive preference threshold are selected, enabling intelligent optimization of circular paths. This makes the final determined target circular paths clearer in visual perception, easier to understand and recognize in structure, and more in line with the cognitive logic and usage habits of airborne visual language navigation. By combining UAV dynamic constraints and cognitive saliency maps, cognitive decision points and physical constraint points on the target circular path are identified. Using these points as segmentation boundaries, the target circular path is divided into multiple continuous smaller segments. This allows for refined segmentation and structured organization of the circular route, taking into account both UAV flight performance constraints and key environmental cognitive nodes, thus improving the rationality and feasibility of path planning. Based on the local fusion feature representations of each smaller segment and its corresponding local fusion feature representation, as well as the global fusion feature representation of the entire target circular path, multi-granularity route guidance instructions containing local action guidance and global semantic descriptions are generated. This enables multi-level, multi-scale navigation guidance from micro-operations to macro-semantics, making route guidance more accurate, clear, and easier to understand and execute. This effectively improves the navigation capability and operational stability of the airborne VLN system in circular route scenarios, enhancing the overall reliability and environmental adaptability of the navigation system. Attached Figure Description
[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0049] Figure 1 This is a flowchart of a circular route guidance method for airborne VLNs disclosed in an embodiment of this application;
[0050] Figure 2 This is a schematic diagram of a circular route guidance device for airborne VLNs disclosed in an embodiment of this application;
[0051] Figure 3 This is a hardware structure block diagram of a ring-shaped route guidance device for airborne VLN disclosed in an embodiment of this application. Detailed Implementation
[0052] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0053] This application can be used in a wide variety of general-purpose or special-purpose computing device environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor devices, distributed computing environments including any of the above devices, etc.
[0054] The following section introduces the solution proposed in this application. The technical solution is as follows, and details are provided below.
[0055] Figure 1 This is a flowchart of a circular route guidance method for airborne VLNs disclosed in an embodiment of this application.
[0056] like Figure 1 As shown, the method may include:
[0057] Step S1: Obtain satellite imagery and vector map of the target area. Using the geometric structure of the vector map as a constraint, map and align the visual features of the satellite imagery to the geometric feature space of the vector map to generate a fused feature representation.
[0058] Specifically, high-resolution satellite imagery of the target area is acquired using satellite remote sensing technology, ensuring clear coverage of the entire geographic scene. Simultaneously, a vector map of the target area is retrieved from a geographic information database. This vector map contains precise geometric structure information of geographic elements such as terrain, roads, and boundaries. Using the geometric structure of the vector map as a spatial constraint benchmark, a feature mapping algorithm is employed to project and align the visual features extracted from the satellite imagery (including texture features, grayscale features, and semantic features) onto the geometric feature space of the vector map. Through coordinate calibration and deviation correction, spatial misalignment between visual and geometric features is eliminated. Ultimately, a fused feature representation is generated that combines the visual and semantic information of the satellite imagery with the geometric spatial information of the vector map, providing unified and accurate feature support for subsequent candidate loop path extraction, path analysis, and guidance instruction generation.
[0059] The method of mapping and aligning the visual features of the satellite imagery to the geometric feature space of the vector map includes: ① In sparse data regions, supplementary image data consistent with the distribution of real images is synthesized based on generative adversarial methods to compensate for feature loss caused by data sparsity and ensure the integrity of the mapping alignment; ② Through contrastive learning, the geometric topology of the vector map is used as positive sample anchor points to maximize the correlation between image blocks and vector structures at the same location in the feature space, realizing the semantic mapping of visual features to geometric feature space. Through the combination of the above two methods, and after subsequent processing such as coordinate calibration and deviation correction, the spatial misalignment between visual features and geometric features is eliminated, and finally a fused feature representation that combines the visual semantic information of satellite imagery and the geometric spatial information of vector map is generated.
[0060] Step S2: Extract closed boundary elements from the vector map as candidate circular paths. Based on the fusion feature representation, analyze the saliency distribution and structural comprehensibility of each candidate circular path at the visual perception level, generate a cognitive saliency map and a path comprehensibility score, and select target circular paths that meet the preset cognitive preference threshold based on the path comprehensibility score.
[0061] Specifically, a geographic feature extraction algorithm is used to traverse all geographic features in the vector map, filtering out boundary features with complete closed shapes (such as regional boundaries, ring road boundaries, etc.), and all of these are selected as candidate ring paths to ensure the diversity and completeness of candidate paths. Based on the fusion feature representation generated in step S1, a visual perception analysis algorithm is used to analyze the visual saliency distribution of each candidate ring path in the target scene, identify visual prominence points on the path, and analyze the structural regularity and spatial coherence of the path to determine the structural intelligibility of each candidate path. Based on the analysis results, a cognitive saliency map that can intuitively reflect the visual prominence of the path is constructed, and a path intelligibility score for each candidate ring path is calculated using a quantitative scoring model. This score is compared with a preset cognitive preference threshold, and candidate ring paths with scores not lower than the threshold are selected as target ring paths for subsequent processing.
[0062] Step S3: Combining the UAV dynamic constraints with the cognitive saliency map, identify the cognitive decision points and physical constraint points on the target circular path, and use the cognitive decision points and physical constraint points as segmentation boundaries to divide the target circular path into multiple continuous small segments.
[0063] Specifically, the dynamic constraints of the UAV are first defined, including core parameters such as maximum flight speed, turning radius, and endurance, which serve as the physical basis for path segmentation. Simultaneously, combined with the cognitive saliency map generated in step S2, nodes with high visual saliency on the target circular path that require visual recognition and decision-making by the UAV are identified as cognitive decision points. Based on the UAV's dynamic constraints, nodes on the target circular path that exceed the UAV's dynamic limits and require flight attitude adjustments are identified as physical constraint points. The identified cognitive decision points and physical constraint points are used together as segmentation boundaries. A segmentation algorithm is employed to divide the target circular path into multiple sequentially connected, appropriately long continuous small segments, achieving refined and structured control of the target circular path, adapting to the UAV's flight characteristics and visual navigation requirements.
[0064] Step S4: Based on each of the small road segments and their corresponding local fusion feature representations, as well as the global fusion feature representation of the entire target loop path, generate a multi-granularity route guidance instruction that includes local action guidance and global semantic description.
[0065] Specifically, for each small road segment divided in step S3, the local fusion feature representation of the corresponding road segment is extracted from the fusion feature representation generated in step S1. Combined with the geometric shape, visual features, and UAV dynamic constraints of the small road segment, local action guidance adapted to the flight of that road segment is generated, including specific operational commands such as flight speed, turning angle, and altitude adjustment. Simultaneously, the global fusion feature representation of the entire target loop path is extracted. Combined with global information such as the overall direction and coverage of the path, a global semantic description is generated to help understand the entire route planning logic, clarifying the overall purpose and navigation objective of the loop path. Integrating the local action guidance with the global semantic description forms multi-granularity route guidance commands that combine micro-level operational details with macro-level semantic guidance. Through feature fusion of satellite imagery and vector maps, cognitively guided target loop path selection, constraint-oriented path segmentation, and multi-granularity guidance command generation, accurate guidance for the aerial VLN loop path is achieved, improving the stability, reliability, and environmental adaptability of the navigation system.
[0066] As can be seen from the above technical solutions, the circular route guidance method and related equipment for airborne VLN provided in this application acquires satellite imagery and vector maps of the target area. Using the geometric structure of the vector map as a constraint, the visual features of the satellite imagery are mapped and aligned to the geometric feature space of the vector map to generate a fused feature representation. This organically unifies visual semantic information and geometric spatial information, constructing an environmental representation that combines geometric accuracy and visual richness, providing more comprehensive and reliable data support for subsequent path analysis and route guidance. By extracting closed boundary elements from the vector map as candidate circular paths, and analyzing the saliency distribution and structural comprehensibility of each candidate circular path at the visual perception level based on the fused feature representation, a cognitive saliency map and a path comprehensibility score are generated. Based on the path comprehensibility score, target circular paths that meet the preset cognitive preference threshold are selected, enabling intelligent optimization of circular paths. This makes the final determined target circular paths clearer in visual perception, easier to understand and recognize in structure, and more in line with the cognitive logic and usage habits of airborne visual language navigation. By combining UAV dynamic constraints and cognitive saliency maps, cognitive decision points and physical constraint points on the target circular path are identified. Using these points as segmentation boundaries, the target circular path is divided into multiple continuous smaller segments. This allows for refined segmentation and structured organization of the circular route, taking into account both UAV flight performance constraints and key environmental cognitive nodes, thus improving the rationality and feasibility of path planning. Based on the local fusion feature representations of each smaller segment and its corresponding local fusion feature representation, as well as the global fusion feature representation of the entire target circular path, multi-granularity route guidance instructions containing local action guidance and global semantic descriptions are generated. This enables multi-level, multi-scale navigation guidance from micro-operations to macro-semantics, making route guidance more accurate, clear, and easier to understand and execute. This effectively improves the navigation capability and operational stability of the airborne VLN system in circular route scenarios, enhancing the overall reliability and environmental adaptability of the navigation system.
[0067] Furthermore, after generating the multi-granularity route guidance instructions, the following is also included:
[0068] ① Construct a UAV dynamics simulation environment, parse the generated multi-granularity flight path guidance instructions into a sequence of control instructions and execute them in the simulation environment, and collect the spatiotemporal deviation between the actual flight trajectory and the planned path;
[0069] ② Based on the spatiotemporal deviation, iteratively optimize the instruction generation strategy until the deviation converges to a preset threshold, and output the final executable route guidance instruction.
[0070] Specifically, based on the actual power parameters, flight characteristics, and geographical environment parameters of the target area, a UAV dynamics simulation environment highly consistent with the real flight scenario is constructed to ensure the accuracy and reliability of the simulation results. The aforementioned multi-granularity flight path guidance instructions are converted into a sequence of control instructions that the UAV can recognize and execute through an instruction parsing module. This sequence of control instructions is then input into the constructed dynamics simulation environment to control the simulated UAV to execute flight tasks according to the instructions. Simultaneously, the actual flight trajectory data of the UAV during the simulation is collected and compared with the preset planned path to calculate the spatiotemporal deviation between the two, including key parameters such as position deviation and time synchronization deviation. Based on the collected spatiotemporal deviation data, the causes of the deviations are analyzed, and the generation strategy of the multi-granularity flight path guidance instructions is adjusted accordingly, including parameter optimization of local action guidance and adjustment of the matching degree between global semantic description and local instructions. Through multiple iterations of optimization, the spatiotemporal deviation is until it converges within the preset deviation threshold range. At this point, the optimized final executable flight path guidance instructions are output to ensure that the UAV can accurately follow the planned path during actual flight, further improving the reliability and accuracy of aerial VLN circular flight path guidance.
[0071] In some embodiments of this application, a multi-index comprehensive screening process may be added to achieve comprehensive and accurate screening of candidate circular paths, taking into account the path's visual adaptability, morphological rationality, and the executableness of subsequent UAV action commands, so as to ensure that the selected target circular path can better adapt to the aerial VLN navigation scenario and the flight characteristics of the UAV.
[0072] The process of filtering to obtain the target circular path also includes:
[0073] Visual and morphological feature indicators of candidate circular paths are extracted, and the target circular path is obtained by comprehensive screening based on the visual feature indicators, the morphological feature indicators, the cognitive saliency map and the path comprehensibility score.
[0074] The visual feature index includes at least one of the following: internal homogeneity index, boundary sharpness index, and landmark availability index. The internal homogeneity index is used to measure the consistency of texture and color within the area surrounded by the path. The boundary sharpness index is used to evaluate the clarity of the path boundary in the image. The landmark availability index is used to identify significant landmark elements present around the path.
[0075] The morphological feature index includes at least one of the path length index, shape index index, and length ratio index, wherein the path length index is used to measure the overall length of the path, the shape index index is used to characterize the geometric complexity and morphological features of the path, and the angle-to-length ratio index is used to characterize the relationship between the path turning angle and the segment length to determine the executability of the action command.
[0076] Specifically, for all candidate loop paths extracted in the previous stage, combined with the fusion feature representation generated in step S1, visual feature indicators and morphological feature indicators corresponding to each candidate path are extracted respectively: For visual feature indicators, the texture similarity and color variance of the area surrounded by the path are calculated using an image texture analysis algorithm to obtain the internal homogeneity index; an edge detection algorithm is used to identify the path boundary, and the boundary sharpness index is calculated using the boundary pixel grayscale difference and edge continuity; a landmark recognition model is used to traverse the area surrounding the path to count the number and clarity of significant landmarks (such as buildings, topographical markers, etc.) to determine the landmark usability index. For morphological feature indicators, the total length of the candidate path is directly measured based on the geometric data of the vector map to obtain the path length index; the shape index is calculated by the ratio of the path perimeter to the area, and the larger the ratio, the higher the geometric complexity of the path; the turning angles and corresponding segment lengths of each segment of the path are counted, and the ratio of the turning angle to the segment length is calculated to obtain the angle-to-length ratio index. When the ratio is within a preset reasonable range, it indicates that the path turning action is highly executable.
[0077] Subsequently, the extracted visual and morphological feature indicators were standardized to eliminate dimensional differences between different indicators and ensure comprehensive comparison of each indicator. Simultaneously, the cognitive saliency map was quantified, converting the visual saliency distribution into a quantitative score, maintaining a consistent dimension with the previously obtained path comprehensibility score. Finally, a comprehensive screening model was constructed. Based on the actual needs of airborne VLN navigation, preset weights were assigned to visual feature indicators, morphological feature indicators, cognitive saliency quantitative scores, and path comprehensibility scores. The comprehensive score of each candidate circular path was calculated, and candidate paths with comprehensive scores higher than preset thresholds and meeting all corresponding requirements were selected as the final target circular paths. This ensures that the target circular paths possess good visual perceptibility and structural comprehensibility, as well as reasonable morphological features and strong action command executability, providing a reliable foundation for subsequent route segmentation and guidance instruction generation.
[0078] Building upon this, to further enhance the flight adaptability of the target circular path and ensure that the selected path fully matches the UAV's flight capabilities, this application may also include:
[0079] The visual feature indicators, morphological feature indicators, cognitive saliency map, path comprehensibility score, and UAV dynamic constraints are weighted and fused to generate a flightability score for each candidate circular path;
[0080] The candidate loop paths are sorted according to the flightability score, and the candidate loop paths whose scores meet the preset threshold are selected as the target loop paths.
[0081] Specifically, the dynamic constraints of the UAV are first quantified. Combining core parameters such as maximum flight speed, turning radius, and endurance, a dynamic constraint quantification index is constructed to ensure it maintains a consistent dimension with visual and morphological feature indices. Then, based on the actual application scenarios of aerial VLN navigation, different preset weights are assigned to the visual feature index, morphological feature index, cognitive saliency quantification score, path comprehensibility score, and UAV dynamic constraint quantification index. The weight of the UAV dynamic constraint can be appropriately increased according to flight safety requirements to prioritize path flyability. A weighted fusion algorithm is used to fuse the above indicators and scores to obtain a flyability score for each candidate circular path. A higher score indicates stronger flight adaptability, safety, and reliability of the candidate path. Then, all candidate circular paths are sorted from highest to lowest flyability score. Candidate paths with scores higher than a preset flyability threshold are selected as the final target circular path, further mitigating flight failures caused by path mismatch with UAV dynamics and improving the safety and feasibility of aerial VLN circular route guidance.
[0082] In some embodiments of this application, the process of step S3, which involves combining the UAV dynamic constraints with the cognitive saliency map to identify cognitive decision points and physical constraint points on the target circular path, and using the cognitive decision points and physical constraint points as segmentation boundaries to divide the target circular path into multiple continuous small segments, is described. Specifically, this may include:
[0083] ① Perform geometric simplification on the target circular path and extract key corner points on the path;
[0084] ② The target circular path is initially segmented using the key corner points to obtain an initial set of road segments;
[0085] ③ Perform peak detection on the cognitive saliency map to identify the path location corresponding to the local maximum point of attention weight, and use it as the cognitive decision point;
[0086] ④ Based on the preset minimum turning radius constraint of the UAV, calculate the radius of curvature of each point on the path, and mark the positions with a radius of curvature less than the minimum turning radius as physical constraint points;
[0087] ⑤ Using the cognitive decision points and the physical constraint points as optimization boundaries, the initial road segment set is divided into two parts. Points where the cognitive decision points and physical constraint points overlap or the distance between them is less than a preset threshold will be used as mandatory segmentation points for the second division.
[0088] ⑥ For road segments whose length exceeds the preset threshold after secondary division, they are uniformly divided at a fixed step size to form the final set of road segments.
[0089] Specifically, the target circular path is first geometrically simplified to extract key corner points. The Douglas-Peucker algorithm is used for geometric simplification, and the specific operation is as follows: a preset simplification threshold is set, and all sampling points of the target circular path are traversed. Taking the starting and ending points of the path as initial vertices, the sampling point with the largest distance between the lines connecting the two points is found. If the maximum distance is greater than the preset simplification threshold, the sampling point is added as a new vertex to the simplified polyline. The above process is repeated until the distance from all sampling points to the corresponding polyline segment is no greater than the preset simplification threshold, at which point the simplification operation is stopped. Finally, all vertices of the simplified polyline are extracted as key corner points. This algorithm can remove redundant sampling points while preserving the core geometric shape of the target circular path, accurately extracting key positions of path turning and direction changes, and providing a reliable basis for subsequent initial segmentation.
[0090] Then, using the extracted key corner points as the boundaries of the initial segmentation, the target loop path is split along each key corner point to obtain an initial set of road segments composed of multiple continuous line segments. Each initial road segment corresponds to a path segment between key corner points, realizing the initial structured division of the target loop path and simplifying the complexity of subsequent optimization processing.
[0091] Peak detection is performed on the generated cognitive saliency map. A local maximum detection algorithm is used to traverse the attention weight values corresponding to all path positions in the cognitive saliency map and identify the positions where the attention weight is at a local maximum. These positions correspond to the areas with the highest visual saliency on the path, which require the UAV to focus on and make visual recognition and navigation decisions. These positions are determined as cognitive decision points to ensure that subsequent segmentation can meet the visual cognition requirements and facilitate the UAV to accurately match visual information and navigation commands.
[0092] First, define the preset minimum turning radius constraint for the UAV. This constraint is determined by the UAV's own power performance and flight control system parameters. Based on the geometric coordinate data of the target circular path, use a curvature calculation algorithm to calculate the curvature radius of each sampling point on the path. Compare the calculated curvature radius with the preset minimum turning radius of the UAV. If the curvature radius of a certain point is less than the minimum turning radius, it indicates that the path turning angle at that position is too large and exceeds the power limit of the UAV. Mark this position as a physical constraint point to ensure that the subsequent road segment division conforms to the flight capability of the UAV and avoid situations where turning actions cannot be performed.
[0093] The identified cognitive decision points and marked physical constraint points are used together as the optimization boundary for secondary partitioning to further split the initial set of road segments. At the same time, a preset distance threshold is set to determine the distance between the cognitive decision points and the physical constraint points. If the two overlap or the distance is less than the preset threshold, the point is used as a forced segmentation point for secondary partitioning and is prioritized for segmentation. This ensures that such key locations can serve as the boundaries of independent road segments, taking into account both visual cognitive needs and drone dynamic constraints.
[0094] Finally, a preset threshold for the length of road segments is set. This threshold is determined based on the drone's flight accuracy requirements and the efficiency of navigation command execution. All the small road segments obtained after the second division are traversed, and the length of each small road segment is calculated. If the length of a small road segment exceeds the preset threshold, the road segment is uniformly divided according to a preset fixed step size. The length of each segment obtained after division does not exceed the preset threshold, and finally a set of small road segments with reasonable length and coherent connection is formed. This provides a refined path foundation for the generation of subsequent multi-granularity flight guidance commands, ensuring that the drone can accurately execute the flight commands of each path segment.
[0095] In some embodiments of this application, the process of generating multi-granularity route guidance instructions containing local action guidance and global semantic description based on each of the small road segments and their corresponding local fusion feature representations, as well as the global fusion feature representation of the entire target loop path, is described. Specifically, it may include:
[0096] ① For each road segment, based on the road segment and its corresponding local fusion feature representation, analyze the boundary geometric features and surrounding land feature distribution of the road segment, and generate local route guidance describing the navigation mode of the road segment;
[0097] ②Based on the global fusion feature representation of the entire target ring path and surrounding area, analyze the semantic structure and spatial relationships of the overall scene to generate a global semantic description of the overall navigation mission;
[0098] ③ Integrate and structure the local route guidance and the global semantic description to generate multi-granularity route guidance instructions.
[0099] Specifically, for each small road segment obtained in step S3, the fusion feature representation generated in step S1 is called to extract the corresponding local fusion feature representation. This local fusion feature representation contains the geometric structure information of the small road segment and the visual semantic information of the surrounding ground features. Based on this local fusion feature representation, a geometric feature analysis algorithm is used to analyze the boundary geometric features of the small road segment, including key parameters such as the road segment's direction, length, turning angle, and curvature changes. At the same time, a ground feature recognition algorithm is used to analyze the distribution of ground features around the small road segment, identifying key elements affecting navigation such as prominent landmarks, obstacles, and terrain undulations. Combining the above analysis results, a local flight path guide adapted to the small road segment is generated. This guide is presented in the form of precise action commands, clearly describing the UAV's navigation mode in this road segment, including specific operational requirements such as flight speed, flight altitude, turning timing, turning angle, and whether visual recognition calibration is required. This ensures that the UAV can accurately execute the navigation task of this road segment and adapt to the geometric features of the road segment and the surrounding environment.
[0100] Subsequently, a global fusion feature representation is extracted for the entire target loop path and the surrounding area. This global fusion feature representation encompasses macroscopic information such as the overall geometry of the target loop path, the distribution of ground features in the surrounding area, and global visual semantics. Based on this global fusion feature representation, a semantic analysis algorithm is used to parse the semantic structure of the overall scene, clarifying the overall purpose of the loop path (such as geographic surveying, disaster patrol, regional monitoring, etc.), its coverage area, and core navigation objectives. Simultaneously, the spatial relationship between the target loop path and the surrounding area's ground features and terrain is analyzed to clarify the overall direction of the path, the distribution of key nodes, and the global navigation logic. Combining the above analysis results, a global semantic description is generated. This description, presented in a concise and clear semantic statement, explains the navigation mission, overall planning logic, and key considerations of the entire loop path, helping the UAV understand the global navigation objective and achieve coordinated unity between local actions and global tasks.
[0101] Finally, the generated local route guidance and global semantic description are semantically integrated and structured: First, the local route guidance is categorized and sorted according to the order of the sub-segments to ensure the coherence and logic of the local guidance; simultaneously, the global semantic description is refined to extract core semantic information, enabling it to correspond with each local route guidance. Then, a semantic fusion algorithm is used, using the global semantic description as the overall guidance framework, embedding each local route guidance into its corresponding position within the framework, achieving precise association between global semantics and local actions; furthermore, the integrated content is structured, using a unified command format to clearly distinguish between the global semantic description and each local route guidance, and labeling the sub-segment numbers corresponding to each local guidance, facilitating rapid identification, parsing, and execution by the UAV. Through the above processing, a multi-granularity route guidance command containing local action guidance and global semantic description is finally generated, providing the UAV with precise segmented flight operation guidance and helping the UAV grasp the overall navigation mission, improving the accuracy, reliability, and executability of aerial VLN navigation.
[0102] The following describes a circular route guidance device for airborne VLNs provided in an embodiment of this application. The circular route guidance device for airborne VLNs described below can be referred to in correspondence with the circular route guidance method for airborne VLNs described above.
[0103] See Figure 2 , Figure 2 This is a schematic diagram of a ring-shaped flight path guidance device for airborne VLNs disclosed in an embodiment of this application.
[0104] like Figure 2 As shown, the circular route guidance device for airborne VLNs may include:
[0105] The data fusion module is used to acquire satellite imagery and vector maps of the target area, and using the geometric structure of the vector map as a constraint, to map and align the visual features of the satellite imagery to the geometric feature space of the vector map to generate a fused feature representation.
[0106] The path filtering module is used to extract closed boundary elements from the vector map as candidate circular paths, analyze the saliency distribution and structural comprehensibility of each candidate circular path at the visual perception level based on the fused feature representation, generate a cognitive saliency map and a path comprehensibility score, and filter out target circular paths that meet the preset cognitive preference threshold based on the path comprehensibility score.
[0107] The path segmentation module is used to combine the UAV dynamic constraints with the cognitive saliency map to identify the cognitive decision points and physical constraint points on the target circular path, and to divide the target circular path into multiple continuous small segments using the cognitive decision points and physical constraint points as segmentation boundaries.
[0108] The instruction generation module is used to generate multi-granularity route guidance instructions that include local action guidance and global semantic description based on each of the small road segments and their corresponding local fusion feature representations, as well as the global fusion feature representation of the entire target loop path.
[0109] As can be seen from the above technical solutions, the circular route guidance method and related equipment for airborne VLN provided in this application acquires satellite imagery and vector maps of the target area. Using the geometric structure of the vector map as a constraint, the visual features of the satellite imagery are mapped and aligned to the geometric feature space of the vector map to generate a fused feature representation. This organically unifies visual semantic information and geometric spatial information, constructing an environmental representation that combines geometric accuracy and visual richness, providing more comprehensive and reliable data support for subsequent path analysis and route guidance. By extracting closed boundary elements from the vector map as candidate circular paths, and analyzing the saliency distribution and structural comprehensibility of each candidate circular path at the visual perception level based on the fused feature representation, a cognitive saliency map and a path comprehensibility score are generated. Based on the path comprehensibility score, target circular paths that meet the preset cognitive preference threshold are selected, enabling intelligent optimization of circular paths. This makes the final determined target circular paths clearer in visual perception, easier to understand and recognize in structure, and more in line with the cognitive logic and usage habits of airborne visual language navigation. By combining UAV dynamic constraints and cognitive saliency maps, cognitive decision points and physical constraint points on the target circular path are identified. Using these points as segmentation boundaries, the target circular path is divided into multiple continuous smaller segments. This allows for refined segmentation and structured organization of the circular route, taking into account both UAV flight performance constraints and key environmental cognitive nodes, thus improving the rationality and feasibility of path planning. Based on the local fusion feature representations of each smaller segment and its corresponding local fusion feature representation, as well as the global fusion feature representation of the entire target circular path, multi-granularity route guidance instructions containing local action guidance and global semantic descriptions are generated. This enables multi-level, multi-scale navigation guidance from micro-operations to macro-semantics, making route guidance more accurate, clear, and easier to understand and execute. This effectively improves the navigation capability and operational stability of the airborne VLN system in circular route scenarios, enhancing the overall reliability and environmental adaptability of the navigation system.
[0110] The circular route guidance device for airborne VLNs provided in this application embodiment can be applied to circular route guidance equipment for airborne VLNs. Figure 3The hardware structure block diagram of a ring-shaped route guidance device for airborne VLNs is shown, with reference to... Figure 3 The hardware structure of a ring-shaped route guidance device for airborne VLNs may include: at least one processor 1, at least one communication interface 2, at least one memory 3, and at least one communication bus 4.
[0111] In this embodiment of the application, the number of processor 1, communication interface 2, memory 3, and communication bus 4 is at least one, and processor 1, communication interface 2, and memory 3 communicate with each other through communication bus 4;
[0112] Processor 1 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.
[0113] Memory 3 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device;
[0114] The memory stores a program, which the processor can call. The program is used for:
[0115] Acquire satellite imagery and vector map of the target area. Using the geometric structure of the vector map as a constraint, map and align the visual features of the satellite imagery to the geometric feature space of the vector map to generate a fused feature representation.
[0116] Closed boundary elements are extracted from the vector map as candidate circular paths. Based on the fusion feature representation, the saliency distribution and structural comprehensibility of each candidate circular path at the visual perception level are analyzed to generate a cognitive saliency map and a path comprehensibility score. Target circular paths that meet the preset cognitive preference threshold are then selected based on the path comprehensibility score.
[0117] By combining the dynamic constraints of the UAV with the cognitive saliency map, the cognitive decision points and physical constraint points on the target circular path are identified, and the target circular path is divided into multiple continuous small segments using the cognitive decision points and the physical constraint points as segmentation boundaries.
[0118] Based on the local fusion feature representations of each of the aforementioned road segments and their corresponding local fusion feature representations, as well as the global fusion feature representations of the entire target loop path, a multi-granularity route guidance instruction containing local action guidance and global semantic description is generated.
[0119] Optionally, the refined and extended functions of the program can be referred to the above description.
[0120] This application embodiment also provides a readable storage medium that can store a program suitable for execution by a processor, the program being used for:
[0121] Acquire satellite imagery and vector map of the target area. Using the geometric structure of the vector map as a constraint, map and align the visual features of the satellite imagery to the geometric feature space of the vector map to generate a fused feature representation.
[0122] Closed boundary elements are extracted from the vector map as candidate circular paths. Based on the fusion feature representation, the saliency distribution and structural comprehensibility of each candidate circular path at the visual perception level are analyzed to generate a cognitive saliency map and a path comprehensibility score. Target circular paths that meet the preset cognitive preference threshold are then selected based on the path comprehensibility score.
[0123] By combining the dynamic constraints of the UAV with the cognitive saliency map, the cognitive decision points and physical constraint points on the target circular path are identified, and the target circular path is divided into multiple continuous small segments using the cognitive decision points and the physical constraint points as segmentation boundaries.
[0124] Based on the local fusion feature representations of each of the aforementioned road segments and their corresponding local fusion feature representations, as well as the global fusion feature representations of the entire target loop path, a multi-granularity route guidance instruction containing local action guidance and global semantic description is generated.
[0125] Optionally, the refined and extended functions of the program can be referred to the above description.
[0126] This application also provides a computer program product, including a computer program, wherein the computer program is executed by a processor using the following method:
[0127] Acquire satellite imagery and vector map of the target area. Using the geometric structure of the vector map as a constraint, map and align the visual features of the satellite imagery to the geometric feature space of the vector map to generate a fused feature representation.
[0128] Closed boundary elements are extracted from the vector map as candidate circular paths. Based on the fusion feature representation, the saliency distribution and structural comprehensibility of each candidate circular path at the visual perception level are analyzed to generate a cognitive saliency map and a path comprehensibility score. Target circular paths that meet the preset cognitive preference threshold are then selected based on the path comprehensibility score.
[0129] By combining the dynamic constraints of the UAV with the cognitive saliency map, the cognitive decision points and physical constraint points on the target circular path are identified, and the target circular path is divided into multiple continuous small segments using the cognitive decision points and the physical constraint points as segmentation boundaries.
[0130] Based on the local fusion feature representations of each of the aforementioned road segments and their corresponding local fusion feature representations, as well as the global fusion feature representations of the entire target loop path, a multi-granularity route guidance instruction containing local action guidance and global semantic description is generated.
[0131] Optionally, the refined and extended functions of the program can be referred to the above description.
[0132] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0133] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0134] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A circular route guidance method for airborne VLNs, characterized in that, include: Acquire satellite imagery and vector map of the target area. Using the geometric structure of the vector map as a constraint, map and align the visual features of the satellite imagery to the geometric feature space of the vector map to generate a fused feature representation. Closed boundary elements are extracted from the vector map as candidate circular paths. Based on the fusion feature representation, the saliency distribution and structural comprehensibility of each candidate circular path at the visual perception level are analyzed to generate a cognitive saliency map and a path comprehensibility score. Target circular paths that meet the preset cognitive preference threshold are then selected based on the path comprehensibility score. By combining the dynamic constraints of the UAV with the cognitive saliency map, the cognitive decision points and physical constraint points on the target circular path are identified, and the target circular path is divided into multiple continuous small segments using the cognitive decision points and the physical constraint points as segmentation boundaries. Based on the local fusion feature representations of each of the aforementioned road segments and their corresponding local fusion feature representations, as well as the global fusion feature representations of the entire target loop path, a multi-granularity route guidance instruction containing local action guidance and global semantic description is generated.
2. The method according to claim 1, characterized in that, The methods for mapping and aligning the visual features of the satellite imagery to the geometric feature space of the vector map include: In data-sparse regions, supplementary image data consistent with the distribution of real images is synthesized based on generative adversarial methods. By using a contrastive learning approach, the geometric topology of the vector map is used as the positive sample anchor point to maximize the correlation between image blocks and vector structures at the same location in the feature space, thereby achieving semantic mapping from visual features to geometric feature space.
3. The method according to claim 1, characterized in that, The process of filtering to obtain the target circular path also includes: Visual and morphological feature indicators of candidate circular paths are extracted, and the target circular path is obtained by comprehensive screening based on the visual feature indicators, the morphological feature indicators, the cognitive saliency map and the path comprehensibility score. The visual feature index includes at least one of the following: internal homogeneity index, boundary sharpness index, and landmark availability index. The internal homogeneity index is used to measure the consistency of texture and color within the area surrounded by the path. The boundary sharpness index is used to evaluate the clarity of the path boundary in the image. The landmark availability index is used to identify significant landmark elements present around the path. The morphological feature index includes at least one of the path length index, shape index index, and length ratio index, wherein the path length index is used to measure the overall length of the path, the shape index index is used to characterize the geometric complexity and morphological features of the path, and the angle-to-length ratio index is used to characterize the relationship between the path turning angle and the segment length to determine the executability of the action command.
4. The method according to claim 3, characterized in that, Also includes: The visual feature indicators, morphological feature indicators, cognitive saliency map, path comprehensibility score, and UAV dynamic constraints are weighted and fused to generate a flightability score for each candidate circular path; The candidate loop paths are sorted according to the flightability score, and the candidate loop paths whose scores meet the preset threshold are selected as the target loop paths.
5. The method according to claim 1, characterized in that, Combining the UAV's dynamic constraints with the cognitive saliency map, cognitive decision points and physical constraint points on the target circular path are identified. Using these cognitive decision points and physical constraint points as segmentation boundaries, the target circular path is divided into multiple continuous small segments, including: The target circular path is geometrically simplified, and key corner points on the path are extracted; The target circular path is initially segmented using the key corner points to obtain an initial set of road segments; Peak detection is performed on the cognitive saliency map to identify the path locations corresponding to the local maxima of attention weights, which are then used as cognitive decision points. Based on the preset minimum turning radius constraint of the UAV, the curvature radius of each point on the path is calculated, and the positions with curvature radii less than the minimum turning radius are marked as physical constraint points; Using the cognitive decision points and physical constraint points as optimization boundaries, the initial road segment set is divided into two parts. Points where the cognitive decision points and physical constraint points overlap or are less than a preset threshold will be used as mandatory segmentation points for the second division. For road segments whose length exceeds a preset threshold after secondary division, they are uniformly divided at a fixed step size to form the final set of road segments.
6. The method according to claim 5, characterized in that, The geometric simplification of the target circular path and the extraction of key corner points on the path include: The Douglas-Peucker algorithm is used to geometrically simplify the target circular path, and the vertices of the simplified polyline are extracted as key corner points.
7. The method according to claim 1, characterized in that, Based on the local fusion feature representations of each of the aforementioned road segments and their corresponding local fusion feature representations, as well as the global fusion feature representation of the entire target loop path, a multi-granularity route guidance instruction containing local action guidance and global semantic description is generated, including: For each road segment, based on the road segment and its corresponding local fusion feature representation, the boundary geometric features and surrounding land cover distribution of the road segment are analyzed to generate local route guidance describing the navigation mode of the road segment; Based on the global fusion feature representation of the entire target ring path and surrounding area, the semantic structure and spatial relationships of the overall scene are analyzed to generate a global semantic description of the overall navigation mission. The local route guidance and the global semantic description are semantically integrated and structured to generate multi-granularity route guidance instructions.
8. The method according to claim 7, characterized in that, After generating the multi-granularity route guidance instructions, the method further includes: A UAV dynamics simulation environment is constructed, and the generated multi-granularity flight path guidance instructions are parsed into a sequence of control instructions and executed in the simulation environment. The spatiotemporal deviation between the actual flight trajectory and the planned path is collected. Based on the spatiotemporal deviation, the instruction generation strategy is iteratively optimized until the deviation converges to a preset threshold, and the final executable route guidance instruction is output.
9. A circular flight path guidance device for aerial vehicle navigation systems (VLNs), characterized in that, include: The data fusion module is used to acquire satellite imagery and vector maps of the target area, and using the geometric structure of the vector map as a constraint, to map and align the visual features of the satellite imagery to the geometric feature space of the vector map to generate a fused feature representation. The path filtering module is used to extract closed boundary elements from the vector map as candidate circular paths, analyze the saliency distribution and structural comprehensibility of each candidate circular path at the visual perception level based on the fused feature representation, generate a cognitive saliency map and a path comprehensibility score, and filter out target circular paths that meet the preset cognitive preference threshold based on the path comprehensibility score. The path segmentation module is used to combine the UAV dynamic constraints with the cognitive saliency map to identify the cognitive decision points and physical constraint points on the target circular path, and to divide the target circular path into multiple continuous small segments using the cognitive decision points and physical constraint points as segmentation boundaries. The instruction generation module is used to generate multi-granularity route guidance instructions that include local action guidance and global semantic description based on each of the small road segments and their corresponding local fusion feature representations, as well as the global fusion feature representation of the entire target loop path.
10. A circular route guidance device for airborne VLNs, characterized in that, Including memory and processor; The memory is used to store programs; The processor is configured to execute the program to implement the various steps of the circular route guidance method for airborne VLNs as described in any one of claims 1-8.
11. A readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the various steps of the circular route guidance method for airborne VLNs as described in any one of claims 1-8.
12. A computer program product, comprising a computer program, characterized in that, The computer program, when run by a processor, performs the various steps of the circular route guidance method for airborne VLNs as described in any one of claims 1-8.