A method, equipment and medium for low-altitude flight path mapping for unmanned aerial vehicles (UAVs)

Abstract

This invention relates to the field of UAV route planning technology, and more particularly to a method, device, and medium for drawing low-altitude routes for UAVs. This invention uses real-world 3D as the basic constraint for 3D planning and design of low-altitude routes. After identification through a semantic segmentation model, each 3D entity is assigned attribute information including flight constraint attributes and differentiated safety margin compensation parameters. The 3D entities establish avoidance zones based on their attribute information. Each unit segment of the avoidance zone and the route is 3D encoded using BeiDou 3D grid position codes. Collision detection is performed on the flight path by taking the intersection, quickly determining route risks. For unit segments with intersection, altitude adjustments and / or detour adjustments are made based on the semantic and attribute information of the 3D entities. This significantly reduces computational complexity and substantially improves route drawing efficiency and low-altitude route utilization.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) route planning technology, and in particular to a method, equipment, and medium for drawing low-altitude routes for UAVs. Background Technology

[0002] Low-altitude flight paths for unmanned aerial vehicles (UAVs) refer to low-altitude flight routes planned for UAVs operating in low-altitude areas (airspace with a true altitude below 1000 meters). The aim is to ensure the safety, orderliness, and efficiency of UAVs during flight, while reducing potential conflicts with other air traffic (such as manned aircraft) and ground activities, and optimizing the utilization of airspace resources.

[0003] Mapping low-altitude flight paths for unmanned aerial vehicles (UAVs) involves multiple aspects, including terrain analysis, obstacle identification, flight path planning, and support from communication and navigation technologies. Currently, low-altitude flight path mapping faces numerous challenges due to the complexity of the low-altitude environment. These challenges include varied terrain, including natural features such as mountains, hills, rivers, and lakes, as well as man-made structures like skyscrapers, bridges, and power lines in cities. This complex terrain and obstacle environment places high demands on UAV flight path planning. High-precision geographic information, including terrain, topography, and obstacle locations, is required to generate safe and efficient low-altitude flight paths.

[0004] Especially in cities, the numerous and complex obstacles such as tall buildings, bridges, and power lines greatly interfere with the mapping of low-altitude flight paths for drones, resulting in low mapping efficiency and difficulty in achieving automated mapping. Summary of the Invention

[0005] To address the problems of numerous and complex obstacles in existing technologies, low efficiency in low-altitude flight path mapping for UAVs, and difficulty in achieving automated flight path mapping, the first objective of this invention is to provide a method for low-altitude flight path mapping for UAVs. This method uses a pre-trained semantic segmentation model to perform entity individualization and semantic segmentation on the oblique photogrammetric 3D MESH data of the area to be planned, identifies the types of each 3D entity and assigns semantic and attribute information, and establishes an avoidance area based on the 3D entity attribute information.

[0006] Based on the subdivision of the BeiDou 3D grid location code, the avoidance area is 3D encoded to generate the grid code sequence of each 3D entity and establish an obstacle 3D entity code library.

[0007] Obtain the coordinates of the starting point and ending point of the route, generate the original continuous route, divide the original continuous route into several continuous unit segments according to the same level of Beidou three-dimensional subdivision grid, construct spatial parameter equations based on the starting point and ending point coordinates of each unit segment, and generate a three-dimensional grid set containing the grid codes of each unit segment.

[0008] The intersection of the 3D mesh set and the obstacle 3D entity code library is taken, and the height and / or detour adjustment of the intersecting unit segments are performed according to the semantic and attribute information of the 3D entities.

[0009] The corrected route is subjected to geometric consistency verification and dynamic smoothing to generate a spatiotemporal trajectory manifold.

[0010] A 3D pipeline model matching the flight safety boundary of a drone is constructed based on the spatiotemporal trajectory manifold.

[0011] Specifically, the attribute information includes flight constraint attributes and differentiated safety margin compensation parameters of the three-dimensional entities. The three-dimensional entities include fixed features, vegetation features, and signal jamming facilities. The differentiated safety margin compensation parameters of the fixed features are smaller than those of the vegetation features, and the differentiated safety margin compensation parameters of the vegetation features are smaller than those of the signal jamming facilities.

[0012] Specifically, the avoidance area is established in a three-dimensional solid based on the constraint height, where the constraint height H is:

[0013] ,

[0014] In the formula, Coordinates of each two-dimensional path node in the UAV's flight path The elevation values ​​on the real-world 3D model are calculated in the MESH triangulation network using the centroid coordinate interpolation method. For differentiated safety margin compensation parameters.

[0015] Specifically, the starting coordinates of any of the aforementioned unit segments End point coordinates The spatial parameter equation for:

[0016] ,

[0017] In the formula, t is a timing parameter. ,

[0018] 3D mesh set It consists of all grid codes that satisfy the following mapping relationship:

[0019] ,

[0020] in, This indicates the floor function. This is the discrete integer index of the mesh in a three-dimensional Cartesian coordinate system. This is the encoding generation function for the BeiDou 3D grid location code. By calculating the intersection points of parameter t at the grid boundary surface, it determines the sequence of all grid indexes traversed by the polyline and outputs the grid code corresponding to each grid cell.

[0021] Specifically, the geometric consistency check is as follows:

[0022] Obtain the set of unit segments and the target direction vector, and calculate the trend vector for each unit segment in the set of unit segments.

[0023] ,

[0024] In the formula, , The coordinates of the starting point of the unit segment. , The coordinates of the end point of the unit segment;

[0025] Calculate the projection relationship between the trend vector and the target vector of a unit segment using the dot product. ,

[0026] ,

[0027] like >0: The angle between the unit segment direction and the target direction is 0 < 90°, which means they are in the same direction.

[0028] like <0: The angle between the unit segment direction and the target direction is 0>90°, which means the opposite direction;

[0029] Reverse the order of the unit segment coordinate sequence.

[0030] Specifically, the pre-trained semantic segmentation model is a PointNet deep neural network model, which uses the color features of texture images to extract two-dimensional semantic information and the geometric features of three-dimensional point clouds to extract spatial structure information; based on the YOLO26 model, synchronous semantic information is extracted from real-time video of UAVs at the same location, and the results of the fusion extraction of real-scene three-dimensional images and point clouds by the PointNet deep neural network model are corrected.

[0031] Specifically, the loss function of the YOLO26 model for:

[0032] ,

[0033] For bounding box regression loss, the full intersection-union loss function is adopted to improve the bounding box localization accuracy and the regression performance for small and dense targets;

[0034] For classification loss, a binary cross-entropy loss with Logits is adopted and a variant of the focus loss function is introduced to alleviate the class imbalance problem;

[0035] To determine the object confidence loss, a binary cross-entropy loss method is used to distinguish between the foreground and background.

[0036] For instance segmentation loss, a composite segmentation loss combining the Dess loss function and the binary cross-entropy loss is used to optimize the region overlap rate and pixel-level classification accuracy, respectively.

[0037] α, β, γ, and δ are the loss weight coefficients for each branch.

[0038] Specifically, the three-dimensional pipeline model is constructed based on the Frene-Cerrey frame, with the spatiotemporal trajectory manifold as the central trajectory reference of the three-dimensional pipeline model; the direction vector and deflection angle of adjacent unit segments are calculated, and a preset heading deflection angle threshold is used to determine whether the heading is jittering. A merging operation is performed on the adjacent nodes of the jittering unit segments; the discrete curvature values ​​of each node of the trajectory are calculated, and for nodes in the spatiotemporal trajectory manifold with abrupt elevation changes and excessive included angles, k-order B-spline curves are used for dynamic smoothing interpolation. By constraining the first and second derivatives of the B-spline curve, the smoothed spatiotemporal trajectory manifold does not exceed the maximum curvature threshold corresponding to the minimum turning radius of the UAV.

[0039] A second objective of the present invention is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0040] A third objective of this invention is to provide a computer-readable storage medium storing a semantic topology descriptor. The semantic topology descriptor is generated using the aforementioned UAV low-altitude flight path drawing method. The topology descriptor has a three-layer decoupled architecture, including a semantic layer, a topology layer, and a dynamics layer. The semantic layer contains three-dimensional entity semantic information and attribute information. The topology layer contains spatiotemporal trajectory manifolds and topological connection relationships between unit segments. The dynamics layer contains UAV physical constraint parameters. A computer constructs a three-dimensional pipeline model based on the semantic topology descriptor.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] 1. This invention uses real-world 3D as the basic constraint for 3D planning and design of low-altitude routes. After identification by a semantic segmentation model, each 3D entity is assigned attribute information including flight constraint attributes and differentiated safety margin compensation parameters. The 3D entity establishes an avoidance zone based on its attribute information. The avoidance zone and each unit segment of the route are 3D encoded based on the BeiDou 3D grid position code. The intersection is used to perform "collision detection" of the flight path, which quickly judges the route risk, greatly reduces the computational complexity, and significantly improves the route drawing efficiency and low-altitude route utilization.

[0043] 2. This invention uses a PointNet deep neural network model to extract two-dimensional semantic information and spatial structure information from real-world 3D models, and combines it with the YOLO26 model to perform synchronous semantic extraction on real-time video of UAVs at the same location. It also corrects the results of the fusion extraction of real-world 3D images and point clouds, thereby improving the accuracy of semantic recognition and segmentation.

[0044] 3. This invention introduces Multidimensional Semantic Topology Descriptor (MSTD) as a lightweight semantic attribute encapsulation protocol for airways, and formulates rules for attributes, geometric logic, etc. for low-altitude airway drawing, which helps to unify standards, promote and popularize the technology, and provide universality for rapid airway drawing using this technology. Attached Figure Description

[0045] Figure 1 This is a flowchart of the present invention;

[0046] Figure 2 This is a roadmap for multimodal feature fusion and semantic segmentation based on real-time video correction in this invention.

[0047] Figure 3 This is a diagram of the MSTD multidimensional semantic topology descriptor architecture of the present invention;

[0048] Figure 4 This is a schematic diagram illustrating the overlay analysis of flight path and physical elements in this invention. Detailed Implementation

[0049] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0050] As shown in the figure, this invention provides a method for drawing low-altitude flight paths for unmanned aerial vehicles (UAVs). Using a pre-trained semantic segmentation model, the method performs entity unitization and semantic segmentation on the oblique photogrammetric 3D mesh data of the area to be planned, identifying the types of each 3D entity and assigning semantic and attribute information. Avoidance zones are established for each 3D entity based on its attribute information. Based on the BeiDou 3D mesh position code, the avoidance zones are 3D encoded, generating mesh code sequences for each 3D entity and establishing a 3D entity code library. The method obtains the starting and ending coordinates of the flight path, generating an original continuous flight path. The original continuous flight path is then divided into several continuous unit segments based on the same-level BeiDou 3D mesh segmentation. Spatial parameter equations are constructed based on the starting and ending coordinates of each unit segment, generating a 3D mesh set containing the mesh codes of each unit segment. The intersection of the 3D mesh set and the obstacle 3D entity code library is taken. For the intersecting unit segments, altitude adjustments and / or detour adjustments are made based on the semantic and attribute information of the 3D entities. The corrected flight path undergoes geometric consistency verification and dynamic smoothing processing to generate a spatiotemporal trajectory manifold. A 3D pipeline model matching the UAV flight safety boundary is constructed based on the spatiotemporal trajectory manifold.

[0051] This invention uses a Multidimensional Semantic Topology Descriptor (MSTD) as a lightweight semantic attribute encapsulation protocol for UAV routes, storing it in a storage medium for efficient and accurate carrying and transmission of geometric, semantic, and dynamic constraint information of the route. The MSTD has a three-layer decoupled architecture, enabling efficient data encapsulation and dynamic injection, including a semantic layer (S-Layer), a topology layer (T-Layer), and a dynamic layer (D-Layer).

[0052] The semantic layer stores the semantic information of each 3D entity. With attribute information, including security level and differentiated safety margin .

[0053] The spatiotemporal trajectory manifold of the route is stored in the topology layer. and the topological connections between unit segments. ,in, , , In three-dimensional coordinates, This corresponds to the timing sequence.

[0054] The dynamics layer stores the physical constraint parameters of the drone, including the drone's minimum turning radius. Maximum rate of ascent, maximum rate of descent, etc.

[0055] MSTD is designed to address the limitations of general geographic data formats (such as GeoJSON), leveraging its customization and high-performance characteristics in the field of UAV low-altitude flight paths. It contributes to the unification of standards and the promotion and popularization of technology, providing universality for rapid flight path mapping.

[0056] During the route drawing process, the system first performs MSTD initialization on the original route. In the subsequent dynamic segmentation and dynamic smoothing steps, the system only needs to perform MSTD initialization on the original route. By redefining and injecting specific dimensions in the data, asynchronous data updates are achieved, avoiding repeated parsing of the entire route file and greatly improving drawing speed and real-time performance.

[0057] S1: Using a pre-trained semantic segmentation model, the oblique photogrammetric 3D MESH data of the planned route area is used to perform entity individualization and semantic segmentation, identify the type of each 3D entity and assign semantic and attribute information, and establish avoidance zones for the 3D entities based on their attribute information.

[0058] Traditional UAV flight is determined by three factors: takeoff point and takeoff altitude, landing point and landing altitude, and flight path altitude. This invention uses a real-world 3D base map and entity element thematic data to perform 3D overlay analysis on UAV flight paths. Through semantic recognition technology, it achieves precise constraints on low-altitude 3D entities and introduces a differentiated safety margin compensation mechanism to improve the precision and safety of flight paths.

[0059] To prepare a real-world 3D base map, this invention preferably uses oblique photogrammetric MESH grid data with a downward view resolution of 15 cm as the base for 3D scene analysis.

[0060] Real-world 3D results are just "a single layer," and machines cannot distinguish between buildings, rivers, and lakes. This invention uses AI to achieve semantic recognition and segmentation through multimodal feature fusion, to identify and classify natural and human entities, and at the same time combine control entities and flight entities to ultimately meet the requirements of UAV flight control.

[0061] This invention acquires a real-world 3D model and its corresponding original texture image and point cloud information. Based on the PointNet deep neural network model, it extracts 2D semantic information using the color features of the texture image and extracts spatial structure information using the geometric features (normal vector, curvature, and elevation) of the 3D point cloud. During the extraction process, this invention performs synchronous semantic extraction on real-time video from a UAV at the same location based on the YOLO26 model, thereby correcting the results of the fusion extraction of the real-world 3D image and the point cloud and improving the accuracy of semantic recognition and segmentation.

[0062] A dedicated training dataset is constructed, comprising real-time drone video sequences whose acquisition locations highly overlap with those of the real-world 3D model, along with corresponding frame-level semantic annotations (including target category labels, bounding box annotations, and pixel-level instance segmentation masks). To improve the model's generalization ability, pre-training can be performed on a large public dataset (such as MS COCO or a custom drone aerial photography semantic dataset), followed by transfer learning and fine-tuning on the specific location matching dataset of this invention. The training process employs an end-to-end supervised learning approach, with AdamW being the preferred optimizer. The learning rate scheduling strategy includes a warm-up phase combined with linear decay to ensure convergence stability. The training batch size is dynamically adjusted based on computing resources, and the number of training epochs is typically 100–300 epochs, with an early stopping mechanism to prevent overfitting.

[0063] The loss function of the YOLO26 model is composed of multi-task losses, as follows:

[0064] .

[0065] (1) For bounding box regression loss, the Complete Intersection-Union-Placement (CIoU) Loss function is adopted to improve the localization accuracy of bounding boxes, especially for small and dense targets such as utility poles and lampposts. It solves the technical problems of gradient vanishing in non-overlapping boxes, inability to distinguish the overlapping of boxes, and large localization deviation of small targets. It provides a precise localization basis for subsequent generation of entity bounding rectangles and construction of vectorized obstacle objects, and meets the accuracy requirements of obstacle avoidance in UAV flight.

[0066] ,

[0067] In the formula, IoU is the crossover-union ratio. B is the predicted bounding box output by the model. These are manually annotated real bounding boxes; Center point of the prediction box Center point of the real frame The squared Euclidean distance is used to penalize the center offset between the two boxes, thus speeding up the model convergence; c is the diagonal length of the smallest bounding rectangle that simultaneously encloses the predicted box and the ground truth box. This is the weighting coefficient for the aspect ratio penalty term, used to balance the distance penalty and the aspect ratio penalty. ; This is a aspect ratio consistency penalty term used to constrain the aspect ratio difference between the predicted bounding box and the ground truth bounding box, thus preventing distortion of the predicted box shape. , , For the width and height of the actual bounding box, This refers to the width and height of the prediction box.

[0068] (2) The classification loss is represented by Binary Cross Entropy with Logits Loss (BCE Loss), and a variant of Focal Loss is introduced. Semantic recognition is a multi-label classification task, where multiple types of entities can coexist in the same scene. The Binary Cross Entropy Loss is used to independently perform binary classification for each category. By modifying the logarithmic operation, the numerical risk of directly calculating log(0) is avoided, ensuring numerical stability during model training and preventing training failure due to gradient vanishing. A variant of Focal Loss is introduced based on the Binary Cross Entropy Loss. Addressing the issues of numerous background samples, few small target samples, and large differences in difficulty between easy and difficult samples in real-world 3D scenes, the loss distribution is optimized through weighted averages of easy and difficult samples and a class balance factor. This doubly suppresses the loss weights of easily distinguishable background samples and amplifies the gradient contribution of difficult and rare targets.

[0069] ,

[0070] Where N is the total number of samples; The model predicts the probability of the true class for positive samples. When negative samples , Use the Sigmoid activation function; β is the focusing factor for easy and difficult samples, used to suppress the loss weight of easy samples and amplify the gradient contribution of difficult samples. Category balancing weights are used to balance the difference in the number of positive and negative samples, alleviating the problem of low proportion of rare category samples; It is the label of the i-th sample.

[0071] (3) The object confidence loss is represented by the BCE loss form, used to distinguish between foreground and background.

[0072] ,

[0073] In the formula, This represents the raw Logits value of the target confidence score output by the model for the r-th anchor box. Let be the actual label of the r-th anchor box, with a value of 1 if it contains a target and a value of 0 if it does not contain a target.

[0074] (4) The instance segmentation loss is represented by a composite segmentation loss combining the Dice loss function and the BCE loss.

[0075] ,

[0076] , The preferred weighting coefficient is the one that can be adjusted. =1.0, =1.5;

[0077] Dice Loss focuses on optimizing the region overlap rate.

[0078] ,

[0079] In the formula, X is the set of foreground pixels predicted by the model; Y is the set of manually labeled real foreground pixels.

[0080] The probabilistic prediction results output by the model are approximated by the sum of element-wise products.

[0081] ,

[0082] In the formula, This represents the probability value of the model output for the pixel at coordinate (i, j) after Sigmoid activation. This represents the true label at pixel coordinates (i, j), with foreground pixels having a value of 1 and background pixels having a value of 0. The smoothing coefficient is preferably set to a value of [value to be filled in]. This is used to avoid calculation errors where the denominator is 0, and to ensure numerical stability;

[0083] BCE loss optimizes pixel-level classification accuracy by breaking down the instance segmentation task into a pixel-by-pixel binary classification task, determining whether each pixel belongs to the foreground or the background.

[0084] ,

[0085] In the formula, For the model coordinates The raw Logits value of the pixel output; coordinates The true label of a pixel, with a value of 1 for foreground pixels and a value of 0 for background pixels.

[0086] α, β, γ, and δ are the loss weight coefficients for each branch, preferably α=1.0, β=1.0, γ=1.0, and δ=0.5.

[0087] Instance segmentation, building upon semantic segmentation, separates different entities belonging to the same category (e.g., two adjacent "residential buildings"), achieving entity semanticization. Independent vectorized obstacle objects are generated based on the bounding rectangle of the identified entities. Simultaneously, attribute information is automatically associated with the individualized objects, including flight constraint attributes and differentiated safety margin compensation parameters for the 3D entities. .

[0088] Natural entities include natural geographical elements such as rivers, lakes, reservoirs, roads, vegetation, and mountains. Natural entities are managed by setting corresponding flight restriction altitudes.

[0089] Human entities include: buildings such as residences, factories, and commercial buildings; obstacles such as high-voltage power line towers, communication towers, wind power towers, and utility poles; and urban infrastructure such as viaducts and elevated roads.

[0090] Controlled entities: airspace above 120 meters in altitude, no-fly zones, restricted airspace and surrounding airspace, airspace for ultra-low-altitude military aviation flights, and controlled airspace.

[0091] Flight entities: Flight entities are aircraft currently flying in the area, including manned aircraft, drones, kites, airships, and other equipment. These entities are processed by adjusting their flight altitude and horizontal distance.

[0092] Flight constraint attributes are inherent control attributes of three-dimensional entities. For example, entities identified as "high-voltage power line towers" are automatically associated with the "electromagnetic interference radius" attribute; entities identified as "water surfaces" are automatically marked as "non-emergency landing zones," etc.

[0093] Differentiated safety margin compensation parameters are set for different semantic entities based on their potential risk to flight, with varying safety margin compensation values. Three-dimensional entities include fixed features, vegetation features, and signal jamming facilities. Fixed features include buildings, ground, and bridges, which pose a lower risk; the preferred differential safety margin compensation parameter is 10 meters. Vegetation features include vegetation and trees, which present risks due to irregular canopies, wind swaying, and uncertainties in plant growth; the preferred differential safety margin compensation parameter is 15 meters. Signal jamming facilities include high-voltage lines and communication towers, which cause electromagnetic interference; the preferred differential safety margin compensation parameter is 25 meters. Through semantic recognition, a uniform and conservative safety distance is avoided for all obstacles, thereby improving the utilization efficiency of low-altitude airspace.

[0094] Based on the attribute information of the 3D entity, an avoidance zone is established. First, the coordinates of each 2D path node of the UAV's flight path are traversed on the real-world 3D base map. By using the centroid coordinate interpolation method, the elevation value of the node in the real-world 3D model under the MESH triangulation was accurately calculated. The data is stored as a three-dimensional array. Based on the flight constraint attributes and differentiated safety margin compensation parameters of the three-dimensional entity, the constraint height of the low-altitude three-dimensional entity is calculated. The constraint value is written into a 3D array. An avoidance area is established on the 3D entity based on the constraint height. The avoidance area can be a rectangular cube extracted outside the 3D entity. The constraint height is used as the height of the cube. A 3D entity library integrating different types of flight constraint obstacles is established.

[0095] For example, avoidance zones are established for areas passing through high-voltage lines based on electromagnetic interference radius and differentiated safety margin compensation parameters; road sections in no-fly zones are marked as no-fly zones and detour strategies are set; and road sections passing through areas such as buildings are established based on their differentiated safety margin compensation parameters.

[0096] Because 3D reality scenes have a timeliness requirement, some areas suffer from outdated data due to long update cycles. To address this timeliness error, this invention introduces a Kalman filter (KF) for dynamic correction, using the state vector... ,in To predict the height for the model, This represents the model error. (Observations) This refers to the real-time vertical height difference between the drone and ground obstacle features. Real-time analysis, using Kalman filtering Iterative updates are performed to achieve dynamic constraints on the route. The filtered model error... When the preset threshold (default value 0.1 meters) is exceeded, the system triggers a partial reconfiguration of the route to ensure the timeliness of the route.

[0097] S2, based on the subdivision of the BeiDou three-dimensional grid position code, performs three-dimensional encoding on the avoidance area, generates the grid code sequence of each three-dimensional entity, and establishes an obstacle three-dimensional entity code library.

[0098] This invention addresses the problem of numerous and complex obstacles in cities by using the Nth level of the BeiDou 3D grid location code to perform 3D encoding on each semantic 3D entity, generating a unique grid code sequence, which is stored in a database. Each 3D entity consists of an array of one or more BeiDou 3D grid location codes, establishing an obstacle 3D entity code library M for easy and rapid indexing.

[0099] The preferred method is to use level 7, with a volumetric dimension of approximately 7.73m × 7.73m × 7.73m. In some embodiments, level 8 may also be used, with a volumetric dimension of approximately 0.97m × 0.97m.

[0100] S3: Obtain the coordinates of the starting point and ending point of the route, generate the original continuous route, divide the original continuous route into several continuous unit segments according to the same level of Beidou three-dimensional subdivision grid, construct spatial parameter equations based on the starting point and ending point coordinates of each unit segment, and generate a three-dimensional grid set containing the grid codes of each unit segment.

[0101] Obtain the coordinates of the starting and ending points of the route to generate the original continuous route. Then, "slice" the original continuous route S (a set of 3D line segments) according to the same-level BeiDou 3D mesh, and set different flight strategies for different "sliced" regions to redefine the MSTD attributes. For example... Figure 4 As shown, the original flight route passes through no-fly zones, high-rise buildings, areas radiated by high-voltage power lines, and mountains, among other physical elements. Different physical elements constrain flight altitude conditions, and a flight route is divided into several continuous unit segments.

[0102] For any cell segment in the three-dimensional space of the original continuous flight path S, the side length of the grid cell at level n of the BeiDou grid code is L. First, it is decomposed into several continuous cell segments. For any given unit segment, its starting point is: The destination is: Construct the spatial parametric equations of the unit segment. :

[0103] ,

[0104] In the formula, t is a timing parameter. ,

[0105] The set of solid meshes that a 3D polyline passes through in 3D space It consists of all grid codes that satisfy the following mapping relationship:

[0106] ,

[0107] in, This indicates the floor function. This is the discrete integer index of the mesh in a three-dimensional Cartesian coordinate system. This is the encoding generation function for the BeiDou 3D grid location code. By calculating the intersection points of parameter t at the grid boundary surface, it determines the sequence of all grid indexes traversed by the polyline and outputs the grid code corresponding to each grid cell.

[0108] S4: Find the intersection of the 3D mesh set and the obstacle 3D entity code library, and adjust the height and / or detour of the intersecting unit segments according to the semantic and attribute information of the 3D entities.

[0109] 3D mesh Intersecting with the obstacle 3D entity code library M, or having no intersection. Sub-elements do not need to have their flight paths adjusted; if there is an intersection, the flight path will be adjusted according to the set. The intersection results with the obstacle 3D entity code library M are used to redefine the text attribute fields of line features. Specifically, based on the obstacle 3D entity category, for entity features such as "buildings", "trees", and "high-voltage power line towers", the height is increased in the Z-axis direction for adjustment, and the text attributes such as "restricted flight altitude", "speed limit", and "heading adjustment" are updated. For no-fly zones, a "no-fly" sign is set and the flight path is adjusted to the periphery of the 3D entity at the current altitude. The system writes the adjusted spatiotemporal trajectory manifold Pi into the T-Layer of MSTD.

[0110] S5 performs geometric consistency verification and dynamic smoothing on the corrected route to generate a spatiotemporal trajectory manifold.

[0111] Iterate through the coordinates of each node in the route, assigning a value to the node height of the route's spatiotemporal trajectory manifold. This is done within the Coordinates string object of the Geometry object. For example, the original path node coordinates are: After assignment, it becomes .

[0112] A flight path often consists of multiple unit segments, represented as a MultiLineString in the Geometry attribute. Multiple flight paths may share some path nodes, leading to inconsistencies in geometric order at merging or connecting points. To address this, each LineString in the MultiLineString needs to be traversed individually, comparing the coordinates of the last node and the first node of adjacent unit segments. If they don't match, the node order of that unit segment is reversed to ensure continuous and consistent path direction, ultimately forming a unified sequence of path coordinates. A reference vector-based geometric direction unification method is employed.

[0113] (1) Variable definition

[0114] Input unit segment set S: a set containing m unit segments L.

[0115] ;

[0116] A single unit segment consists of a series of ordered coordinate points.

[0117] ,

[0118] in, This is the starting point of the unit segment. The endpoint;

[0119] Target direction vector The expected uniform direction (normalized vector), for example: from left to right (i.e., from west to east): From bottom to top (i.e., from south to north): .

[0120] (2) Feature vector extraction

[0121] For each unit segment in the set Calculate its trend vector The trend is determined by the displacement of the endpoint relative to the starting point.

[0122] ,

[0123] In the formula, , The coordinates of the starting point of the unit segment. , The coordinates are the endpoint coordinates of the unit segment.

[0124] (3) Same direction determination

[0125] Calculate the projection relationship between the trend vector and the target vector of a unit segment using the dot product. ,

[0126] ,

[0127] like >0: The angle between the unit segment direction and the target direction is 0 < 90°, which means they are in the same direction.

[0128] like <0: The angle between the unit segment direction and the target direction is 0>90°, which means the opposite direction.

[0129] (4) Correction operation

[0130] Define the modified unit segment If the result is determined to be reversed, the coordinate sequence is then reversed.

[0131] .

[0132] (5) Global sorting

[0133] If the order of the unit segments in the list is standardized (e.g., starting from the leftmost line), then the corrected set S' is sorted. This invention standardizes the "left-to-right" processing, sorting by the x-coordinate of the starting point:

[0134] .

[0135] For areas with gentle elevation changes, Z-value interpolation optimization is used, and linear interpolation is used to supplement the Z-values ​​of path nodes to reduce data redundancy; in areas with abrupt terrain changes, the original measured elevation points are retained to ensure flight safety.

[0136] Based on the redefined 3D coordinate sequence, topological relationship verification is performed, a flight path topology network is constructed, and the presence of abnormal structures such as self-intersections, overlaps, or breaks is checked. Buffer analysis is introduced to assist in corrections when necessary. Finally, the spatiotemporal trajectory manifold of the MSTD's T-Layer is updated to support subsequent 3D flight path rendering.

[0137] S6 constructs a 3D pipeline model that matches the flight safety boundary of UAVs based on the spatiotemporal trajectory manifold.

[0138] The system employs WebGL-based instantiation rendering and shader volume rendering techniques to reconstruct flight paths within a realistic 3D digital twin scene. The system analyzes and extracts the spatiotemporal trajectory manifold as the initial center trajectory for the 3D pipeline model, establishing a 3D pipeline model based on the Frenet-Serret frame to visually represent the flight safety corridor. For path continuity processing, endpoint topological connectivity is ensured by introducing a discrete curvature constraint algorithm. This algorithm monitors abrupt changes in heading in real time by calculating the dot product and cross product of adjacent waypoint vectors. Anisotropic filtering is applied to minor jitter, and for nodes with abrupt elevation changes or large curvature, a method satisfying C... 2 Continuous B-splines (B-Splines) are used for dynamic smoothing interpolation to ensure that the generated trajectory meets the physical constraints of the UAV's minimum turning radius and maximum climb rate. The formula used is as follows:

[0139] (1) In order to "filter out small deflections", a discriminant formula based on vectors is established, and the direction vector of the i-th unit segment is set. for:

[0140] ,

[0141] In the formula, For the (i+1)th unit segment, For the i-th unit segment,

[0142] Then the deflection angle of adjacent unit segments Defined as:

[0143] ,

[0144] Filtering strategy formula: Define the filtering function Set threshold (Default value: 2 degrees)

[0145] ,

[0146] when When this occurs, it is considered a jitter, and a merge operation is performed. At that time, the original route will be retained.

[0147] (2) There are elevation abrupt changes in the spatiotemporal trajectory manifold, i.e., line segments where the trajectory elevation change is greater than the maximum climb rate and maximum descent rate of the UAV, or there is an angle exceeding the limit, i.e., if the discrete curvature of the trajectory exceeds the maximum allowable curvature corresponding to the minimum turning radius, a k-order B-spline curve C(u) is used to generate a B-spline smooth curve that satisfies the dynamic constraints, so as to achieve a smooth transition of "elevation abrupt changes or angle exceeding the limit" and ensure continuous acceleration (continuous) (to improve flight stability)

[0148] ,

[0149] in, For normalized time parameters, These are the control vertices after geometric correction. These are basis functions based on the Cox-de-Boor recursive formula.

[0150] (3) In order to “improve flight stability”, the curvature k(u) of the curve must be less than the maximum curvature allowed by the UAV’s maneuverability. :

[0151] ,

[0152] in, The velocity vector of the drone. Let this be the acceleration vector of the drone. This is the minimum turning radius for a drone as defined in MSTD.

[0153] The smoothed 3D coordinate sequence is written into the MSTD T-Layer as the final spatiotemporal trajectory manifold, and finally outputs a route file that conforms to the MSTD protocol specification.

[0154] This flight path model supports dynamic updates and multi-aircraft collaborative path planning. It can integrate IoT meteorological data in real time, further improving adaptability and safety in complex environments, and providing reliable technical support for UAV operations in areas with weak perception, such as tunnels and canyons.

[0155] Some embodiments also include a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above method.

[0156] Some embodiments also include a computer-readable storage medium storing a semantic topology descriptor, on which a computer constructs a three-dimensional pipeline model based on semantic layer, topology layer, and dynamics layer information in the semantic topology descriptor.

[0157] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention. The above are merely preferred embodiments of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the protection scope of the present invention.

Claims

1. A method for drawing low-altitude flight paths of unmanned aerial vehicles (UAVs), characterized in that, include: Using a pre-trained semantic segmentation model, the oblique photogrammetric 3D MESH data of the planned route area is used to perform entity individualization and semantic segmentation, identify the types of each 3D entity and assign semantic and attribute information, and establish avoidance areas based on the 3D entity attribute information. Based on the subdivision of the BeiDou 3D grid location code, the avoidance area is 3D encoded to generate the grid code sequence of each 3D entity and establish an obstacle 3D entity code library. The origin and destination coordinates of the route are obtained to generate an original continuous route. Based on a corresponding BeiDou 3D mesh, the original continuous route is divided into several continuous unit segments. Spatial parameter equations are constructed based on the origin and destination coordinates of each unit segment, generating a 3D mesh set containing the mesh codes of each unit segment. The starting coordinates of any of the aforementioned unit segments End point coordinates The spatial parameter equation for: ， In the formula, t is a timing parameter. , 3D mesh set It consists of all grid codes that satisfy the following mapping relationship: ， in, This indicates the floor function. This is the discrete integer index of the mesh in a three-dimensional Cartesian coordinate system. This is the encoding generation function for the BeiDou 3D grid location code. By calculating the intersection of parameter t at the grid boundary surface, it determines the sequence of all grid indexes traversed by the polyline and outputs the grid code corresponding to each grid cell. The intersection of the 3D mesh set and the obstacle 3D entity code library is taken, and the height and / or detour adjustment of the intersecting unit segments are performed according to the semantic and attribute information of the 3D entities. The corrected route is subjected to geometric consistency verification and dynamic smoothing to generate a spatiotemporal trajectory manifold. The geometric consistency verification is as follows: Obtain the set of unit segments and the target direction vector, and calculate the trend vector for each unit segment in the set of unit segments. , ， In the formula, , The coordinates of the starting point of the unit segment. , The coordinates of the end point of the unit segment; Calculate the projection relationship between the trend vector and the target vector of a unit segment using the dot product. , ， like >0: The angle between the unit segment direction and the target direction is 0 < 90°, which means they are in the same direction. like <0: The angle between the unit segment direction and the target direction is 0>90°, which means the direction is reversed. Reverse the order of the reversed unit segment coordinate sequence; A 3D pipeline model matching the flight safety boundary of a drone is constructed based on the spatiotemporal trajectory manifold.

2. The method for drawing low-altitude flight paths of unmanned aerial vehicles according to claim 1, characterized in that, The attribute information includes flight constraint attributes and differentiated safety margin compensation parameters of the three-dimensional entities. The three-dimensional entities include fixed features, vegetation features, and signal jamming facilities. The differentiated safety margin compensation parameters of the fixed features are smaller than those of the vegetation features, and the differentiated safety margin compensation parameters of the vegetation features are smaller than those of the signal jamming facilities.

3. The method for drawing low-altitude flight paths of unmanned aerial vehicles according to claim 2, characterized in that, The avoidance zone is established in a three-dimensional solid based on the constraint height H, where the constraint height H is: ， In the formula, Coordinates of each two-dimensional path node in the UAV's flight path The elevation values ​​on the real-world 3D model are calculated in the MESH triangulation network using the centroid coordinate interpolation method. For differentiated safety margin compensation parameters.

4. The method for drawing low-altitude flight paths of unmanned aerial vehicles according to claim 1, characterized in that, The pre-trained semantic segmentation model is a PointNet deep neural network model, which uses the color features of texture images to extract two-dimensional semantic information and the geometric features of three-dimensional point clouds to extract spatial structure information. Based on the YOLO26 model, synchronous semantic information is extracted from the real-time video of the UAV at the corresponding acquisition location of the real-scene three-dimensional model, and the extraction results of the real-scene three-dimensional image and point cloud fusion of the PointNet deep neural network model are corrected.

5. The method for drawing low-altitude flight paths of unmanned aerial vehicles according to claim 4, characterized in that, The loss function of the YOLO26 model for: ， For bounding box regression loss, the full intersection-union loss function is adopted to improve the bounding box localization accuracy and the regression performance for small and dense targets; For classification loss, a binary cross-entropy loss with Logits is adopted and a variant of the focus loss function is introduced to alleviate the class imbalance problem; To determine the object confidence loss, a binary cross-entropy loss method is used to distinguish between the foreground and background. For instance segmentation loss, a composite segmentation loss combining the Dessian loss function and the binary cross-entropy loss is adopted to optimize the region overlap rate and pixel-level classification accuracy, respectively. α, β, γ, and δ are the loss weight coefficients for each branch.

6. The method for drawing low-altitude flight paths of unmanned aerial vehicles according to claim 1, characterized in that, The three-dimensional pipeline model is constructed based on the Frene-Cerrey frame, with the spatiotemporal trajectory manifold as the central trajectory reference of the three-dimensional pipeline model. The direction vector and deflection angle of adjacent unit segments are calculated, and a preset heading deflection angle threshold is used to determine whether the heading is jittering. A merging operation is performed on the adjacent nodes of jittering unit segments. The discrete curvature values ​​of each node of the trajectory are calculated. For nodes with abrupt elevation changes and excessive included angles in the spatiotemporal trajectory manifold, dynamic smoothing interpolation is performed using a k-order B-spline curve. By constraining the first and second derivatives of the B-spline curve, the smoothed spatiotemporal trajectory manifold does not exceed the maximum curvature threshold corresponding to the minimum turning radius of the UAV.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The storage medium stores a semantic topology descriptor, which is generated using the UAV low-altitude flight path drawing method as described in any one of claims 1-6. The topology descriptor has a three-layer decoupled architecture, including a semantic layer, a topology layer, and a dynamics layer. The semantic layer contains three-dimensional entity semantic information and attribute information. The topology layer contains the spatiotemporal trajectory manifold and the topological connection relationship between each unit segment. The dynamics layer contains UAV physical constraint parameters. The computer constructs a three-dimensional pipeline model based on the semantic topology descriptor.

Images