Unmanned aerial vehicle flight environment reconstruction and situation awareness method based on edge computing

Abstract

The present application relates to the technical field of unmanned aerial vehicle intelligent sensing, in particular to a method for reconstructing and perceiving the flight environment of unmanned aerial vehicles based on edge computing, which comprises: an unmanned aerial vehicle carrying a multi-source sensor array synchronously collecting visible light images, depth point clouds and radio frequency signal features, completing spatio-temporal registration through an onboard fusion computing unit, generating registered heterogeneous data blocks and mapping through feature extraction and gridding, and constructing a three-dimensional local environment feature grid centered on the unmanned aerial vehicle. After the grid is transmitted to an edge situation node, multi-source grid splicing and conflict resolution are performed, and a unified wide-area environment model is fused, based on which future situation deduction is carried out and flight instructions are generated. This method eliminates spatio-temporal deviations of heterogeneous data, solves multi-aircraft data splicing conflicts, realizes unified representation of wide-area environment, strengthens the relevance of environment perception, and makes flight instructions conform to the actual situation.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensing technology for unmanned aerial vehicles (UAVs), and in particular to a method for UAV flight environment reconstruction and situational awareness based on edge computing. Background Technology

[0002] Conventional UAV flight situational awareness typically relies on a single sensor to collect environmental data. The onboard unit only completes basic data acquisition and preliminary transmission. Heterogeneous sensing information such as visible light images, depth point clouds, and radio frequency signal characteristics is not synchronously spatiotemporally registered. Environmental information is often represented as scattered two-dimensional data or irregular three-dimensional point clouds, failing to form a standardized three-dimensional local environmental feature mesh centered on the UAV. Edge nodes on the communication base station side only perform data aggregation and forwarding functions, without conducting specialized processing for environmental data from multiple UAVs.

[0003] Under current technology, heterogeneous environmental perception data suffers from spatiotemporal misalignment, making it impossible to form a unified and correlated representation of multiple types of environmental information, and local environmental features lack a regular rasterized presentation format. When environmental data collected by multiple UAVs from different perspectives and at different time series converge at the edge, information conflicts and splicing gaps occur, making it impossible to fuse them into a unified wide-area environmental model. Situation simulation can only be carried out based on local and scattered data, and flight command generation lacks support from the entire environmental domain.

[0004] This invention aims to eliminate the spatiotemporal bias of visible light images, depth point clouds, and radio frequency signal features, complete the feature extraction and rasterization mapping of heterogeneous data after registration, construct a three-dimensional local environment feature grid centered on the UAV, and simultaneously perform splicing and conflict resolution on the three-dimensional local environment feature grids of multiple UAVs from different perspectives and time sequences to form a unified wide-area environment model and carry out situational simulation to generate flight commands. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for UAV flight environment reconstruction and situational awareness based on edge computing.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for UAV flight environment reconstruction and situational awareness based on edge computing, comprising:

[0007] The multi-source sensor array carried by the UAV synchronously collects environmental perception information of the flight airspace, including visible light images, depth point clouds and radio frequency signal characteristics.

[0008] The environmental perception information is input into the airborne fusion computing unit, and the visible light image, depth point cloud and radio frequency signal features are spatiotemporally registered in the airborne fusion computing unit to generate a registered heterogeneous data block.

[0009] Feature extraction and rasterization mapping are performed on the registered heterogeneous data blocks to construct a three-dimensional local environment feature mesh centered on the UAV;

[0010] The three-dimensional local environment feature mesh is transmitted to the edge situation node deployed at the communication base station, where the three-dimensional local environment feature mesh from multiple UAVs is received.

[0011] Inside the edge situation node, multi-source mesh splicing and conflict resolution operations are performed to merge multiple three-dimensional local environmental feature meshes from different perspectives and times into a unified wide-area environmental model.

[0012] Based on the wide-area environment model, future situation simulations are performed to generate UAV flight commands.

[0013] As a further aspect of the present invention, feature extraction and rasterization mapping are performed on the registered heterogeneous data blocks to construct a three-dimensional local environment feature mesh centered on the UAV, including:

[0014] Texture feature sub-blocks, geometric structure sub-blocks, and signal strength sub-blocks are separated from the registered heterogeneous data blocks;

[0015] Edge and corner detection is performed on the texture feature sub-blocks to extract contour features from the image;

[0016] The surface normal and curvature of the geometric structure sub-blocks are estimated and the plane and edge features in the point cloud are extracted.

[0017] The signal source direction is estimated for the signal strength sub-block, and the spatial distribution characteristics of the radio frequency signal are extracted.

[0018] Define a fixed-size three-dimensional voxel space, and map the extracted contour features, plane features, edge features, and spatial distribution features to the corresponding voxels according to their spatial coordinates;

[0019] A comprehensive feature vector is calculated for each voxel filled with features, and all voxels and their corresponding comprehensive feature vectors together constitute the three-dimensional local environment feature mesh.

[0020] As a further aspect of the present invention, within the edge situation node, multi-source mesh stitching and conflict resolution operations are performed to fuse multiple three-dimensional local environment feature meshes from different perspectives and times into a unified wide-area environment model, including:

[0021] Add a spatiotemporal stamp and UAV position and attitude information to each received 3D local environment feature mesh;

[0022] Using the global coordinate system maintained by the edge situation nodes as a reference, all three-dimensional local environment feature meshes are transformed to the global coordinate system;

[0023] In the global coordinate system, overlap detection is performed on voxels describing the same spatial region in different 3D local environment feature meshes;

[0024] For overlapping voxels, their comprehensive feature vectors are compared, and a confidence-weighted fusion algorithm is used to generate new comprehensive feature vectors.

[0025] For voxels in non-overlapping regions, they are directly incorporated into the fusion result, ultimately forming the wide-area environment model that covers a wider spatial range.

[0026] As a further aspect of the present invention, the confidence-weighted fusion algorithm dynamically calculates the confidence weight of voxel features in each grid based on the acquisition sensor accuracy, the distance from the target during acquisition, and the data freshness of each three-dimensional local environmental feature grid.

[0027] As a further aspect of the present invention, generating UAV flight commands based on the wide-area environment model through future situational simulation includes:

[0028] Based on the wide-area environment model, preset semantic understanding rules are invoked to identify and classify the elements in the model, and static terrain structures, dynamic obstacles and potential risk areas are marked.

[0029] Using the wide-area environment model and annotation results, a virtual physics engine is driven to perform short-time domain future situation simulation and predict the state changes of various elements in the environment over a period of time.

[0030] Based on the results of the future situation projection, calculate the airspace passage cost map of the drone swarm at the present and future times;

[0031] By combining the airspace access cost map with the UAV's scheduled flight mission, real-time obstacle avoidance and path replanning instructions are generated for each UAV.

[0032] The real-time obstacle avoidance and path replanning commands are sent to the corresponding UAVs to update the target waypoints and flight strategies in their flight control systems.

[0033] As a further aspect of the present invention, based on the wide-area environment model, preset semantic understanding rules are invoked to identify and classify the elements in the model, specifically including:

[0034] Define a set of semantic categories, which at least include building surfaces, vegetation canopies, moving vehicles, pedestrians, and high-voltage cables;

[0035] A corresponding feature matching template is set for each semantic category. The feature matching template describes the typical combination of texture features, geometric features and spatial distribution features of the corresponding semantic category.

[0036] Traverse each voxel or set of voxels in the wide-area environment model and calculate the similarity between its comprehensive feature vector and the feature matching template of each semantic category.

[0037] Voxels or sets of voxels are classified into the semantic categories with the highest similarity, and the corresponding category labels are assigned to the voxels or sets of voxels.

[0038] As a further aspect of the present invention, the wide-area environment model and annotation results are used to drive a virtual physics engine to perform short-time domain future situation prediction, including:

[0039] Elements classified as dynamic obstacles are extracted from the wide-area environment model, and simplified dynamic models are established for them;

[0040] Static terrain structures are extracted from the wide-area environment model and used as collision boundaries and motion constraints in the virtual physics engine.

[0041] For each dynamic obstacle, set initial values ​​for its motion state based on the current observations and assign it behavioral rules that conform to its semantic category.

[0042] In the virtual physics engine, the physical effects of gravity and wind resistance are simulated, and the collision boundaries and motion constraints are loaded.

[0043] The virtual physics engine is run to calculate the predicted positions and motion states of all dynamic obstacles in multiple consecutive time slices in the future, forming a future situation sequence.

[0044] As a further aspect of the present invention, based on the results of the future situation projection, the airspace passage cost map of the UAV swarm at the current and future times is calculated, including:

[0045] Using the aforementioned wide-area environment model as a spatial basis, it is divided into three-dimensional spatial cells;

[0046] Define a cost calculation function that comprehensively considers the distance between the cell and the static terrain structure, the distance between the cell and the dynamic obstacle at the prediction time, and the signal occlusion degree of the cell.

[0047] The predicted location of the dynamic obstacle at each moment in the future situation sequence is input into the cost calculation function at the corresponding moment.

[0048] For each future time slice, calculate the passage cost of each three-dimensional spatial cell and generate a static and dynamic integrated passage cost distribution map for the corresponding future time slice;

[0049] The sequence of comprehensive traffic cost distribution maps arranged in chronological order is integrated into the aforementioned four-dimensional spatiotemporal airspace traffic cost map.

[0050] As a further aspect of the present invention, by combining the airspace access cost map with the UAV's predetermined flight mission, real-time obstacle avoidance and path replanning instructions applicable to each UAV are generated, including:

[0051] Analyze the planned flight mission of each drone and extract its target point sequence and time constraints;

[0052] Starting from the current location of the UAV and ending at the mission target point, search for a spatiotemporal trajectory from the current time to the target time in the airspace passage cost map.

[0053] The optimization objective of the search is to minimize the total travel cost of all spatiotemporal cells traversed by the entire spatiotemporal trajectory, while simultaneously satisfying the dynamic constraints of the UAV.

[0054] The optimal spatiotemporal trajectory obtained from the search is solved into a series of spatial coordinate points and velocity commands ordered by time, which constitute the real-time obstacle avoidance and path replanning commands.

[0055] As a further aspect of the present invention, the search process employs the spatiotemporal A* algorithm in a four-dimensional spatial travel cost map, with time used as an additional one-dimensional extended cost.

[0056] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0057] The airborne terminal performs synchronous spatiotemporal registration of visible light images, depth point clouds, and radio frequency signal features. Based on the registered heterogeneous data blocks, feature extraction and rasterization mapping are carried out to build a three-dimensional local environment feature grid centered on the UAV. The spatiotemporal deviation of heterogeneous sensing data is eliminated, and multiple types of environmental information form a unified spatiotemporal correlation form. The spatial features and attribute information of the three-dimensional local environment are presented in a regular raster form. The integrity and correlation of local environmental sensing information are enhanced, and the transmission form of local environmental features has standardized and lightweight characteristics.

[0058] Edge situation nodes receive 3D local environmental feature meshes transmitted from multiple UAVs, perform multi-source mesh stitching and conflict resolution operations, and merge multi-view, multi-temporal local meshes into a unified wide-area environment model. Based on the wide-area environment model, future situation simulation is carried out and UAV flight commands are generated. The perspective deviation and temporal conflict between local mesh data collected by different UAVs are resolved. Multi-source local meshes achieve seamless connection of the entire dimension. The overall structure and dynamic characteristics of the wide-area environment form a single unified representation. Situation simulation is carried out based on the full-area environment data, and flight commands are generated to fit the real situation changes of the wide-area environment. Attached Figure Description

[0059] Figure 1 This is a flowchart of the UAV flight environment reconstruction and situational awareness method based on edge computing described in this invention;

[0060] Figure 2 A flowchart for multi-source mesh stitching and conflict resolution operations;

[0061] Figure 3 Heatmap of templates for matching environmental semantic category features;

[0062] Figure 4 The changing trend of multi-time-slice access costs in key areas;

[0063] Figure 5 A comparison chart of the cumulative travel costs for path planning in 3D A and spatiotemporal A. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0065] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0066] See Figure 1The overall implementation scheme of the UAV flight environment reconstruction and situational awareness method based on edge computing provided by this invention is as follows. The method involves a multi-source sensor array carried by the UAV synchronously collecting environmental perception information of the flight airspace. This environmental perception information includes visible light images, depth point clouds, and radio frequency signal features. The collected environmental perception information is input to an onboard fusion computing unit, where the visible light images, depth point clouds, and radio frequency signal features are spatiotemporally registered to generate registered heterogeneous data blocks. Subsequently, the registered heterogeneous data blocks are processed to construct a three-dimensional local environmental feature mesh centered on the UAV. This three-dimensional local environmental feature mesh is transmitted to an edge situational awareness node deployed at a communication base station, where it receives three-dimensional local environmental feature meshes from multiple UAVs. Inside the edge situational awareness node, operations are performed to fuse multiple three-dimensional local environmental feature meshes from different perspectives and times into a unified wide-area environment model. Finally, based on the wide-area environment model, future situational prediction is performed, and UAV flight commands are generated.

[0067] In one embodiment of the present invention, during the construction of a three-dimensional local environment feature mesh, texture feature sub-blocks, geometric structure sub-blocks, and signal intensity sub-blocks are separated from the registered heterogeneous data blocks. Edge and corner detection is performed on the texture feature sub-blocks to extract contour features from the image. Surface normal estimation and curvature calculation are performed on the geometric structure sub-blocks to extract plane and edge features from the point cloud. Signal source direction estimation is performed on the signal intensity sub-blocks to extract the spatial distribution features of the radio frequency signal. A fixed-size three-dimensional voxel space is defined, and the extracted contour features, plane, edge features, and spatial distribution features are mapped to corresponding voxels according to their spatial coordinates. A comprehensive feature vector is calculated for each voxel filled with features, and all voxels and their corresponding comprehensive feature vectors together constitute the three-dimensional local environment feature mesh.

[0068] In practical implementation, a drone flies through an urban street scene. The drone's multi-source sensor array synchronously collects environmental perception information, including visible light images, depth point clouds, and radio frequency (RF) signal features. The visible light images are captured by an onboard camera, the depth point clouds are generated by a lidar sensor, and the RF signal features are obtained by a spectrum analysis module. In the example scenario, the drone flies at an altitude of 50 meters and a speed of 5 meters per second. The sensors collect data at a frequency of 10 frames per second. The collected raw data includes image pixel matrices, point cloud coordinate sets, and signal intensity spectra. This data is input into the onboard fusion computing unit in real time. The onboard fusion computing unit performs spatiotemporal registration on the visible light images, depth point clouds, and RF signal features, generating registered heterogeneous data blocks. The registration operation is achieved through timestamp alignment and coordinate transformation, ensuring that image pixels, point cloud points, and RF sampling points are consistent in time and space, forming heterogeneous data blocks with a unified time reference and spatial reference.

[0069] From the registered heterogeneous data blocks, texture feature sub-blocks, geometric structure sub-blocks, and signal intensity sub-blocks are separated. Texture feature sub-blocks correspond to pixel regions in the visible light image, geometric structure sub-blocks correspond to the 3D coordinates of the depth point cloud, and signal intensity sub-blocks correspond to the feature vectors of the radio frequency (RF) signal. Edge and corner detection is performed on the texture feature sub-blocks to extract contour features from the image. Edge detection uses the Canny operator, and corner detection uses the Harris corner detection algorithm. In the example scene, the edges of building windows and streetlight corners are extracted as contour features from a single frame. Surface normal estimation and curvature calculation are performed on the geometric structure sub-blocks to extract plane and edge features from the point cloud. Principal component analysis is used for surface normal estimation, and curvature calculation is based on local surface fitting. In the example, wall planes and roof edges in the point cloud data are identified as geometric features. Signal source direction estimation is performed on the signal intensity sub-blocks to extract the spatial distribution features of the RF signal. Direction estimation uses a multi-signal classification algorithm. In the example, the direction of arrival of the mobile communication base station signal is identified. Define a fixed-size three-dimensional voxel space, with voxel dimensions of 0.1 m × 0.1 m × 0.1 m. Map the extracted contour features, plane features, edge features, and spatial distribution features to the corresponding voxels according to their spatial coordinates. In the example scene, a voxel may simultaneously contain window edge features from the image, planar features from the point cloud, and specific directional features from the radio frequency signal.

[0070] A comprehensive feature vector is calculated for each voxel filled with features. All voxels and their corresponding comprehensive feature vectors together constitute a 3D local environment feature mesh. The comprehensive feature vector is calculated by weighted fusion of sub-features, as expressed by the formula:

[0071] v 综合 =w t ·f t +w g ·f g +w s ·f s

[0072] Where: v 综合 It is a comprehensive feature vector, f t It is the contour feature vector extracted from the texture feature sub-blocks, f g f is the feature vector of planes and edges extracted from the geometric structure sub-blocks. s w is the spatial distribution feature vector extracted from the signal strength sub-block. t It is the texture feature weight, w g It is the geometric feature weight, w s These are signal feature weights, with weight values ​​preset based on sensor confidence levels. In this example, the lidar has high accuracy, hence w gThe weights are set to 0.5, with the camera and RF sensor weights at 0.3 and 0.2 respectively. In data comparison, when raw heterogeneous data blocks without feature extraction are directly mapped to voxels, the data redundancy within the voxels is high and the dimensions are inconsistent. However, by extracting features and mapping them using this implementation, the comprehensive feature vector within the voxels has a unified dimension of a fixed length, reducing the data volume and improving feature representativeness. For example, in the same urban scene, the original data block size is 10MB, while the generated 3D local environment feature mesh after feature extraction occupies only 2MB. Furthermore, the feature vectors are more conducive to subsequent processing. In some embodiments, the size of the voxel space can be dynamically adjusted according to the UAV's flight altitude; for example, the voxel size is set to 0.05 meters during low-altitude fine scanning. Optionally, a deep learning model can be used instead of a traditional algorithm for feature extraction to improve feature robustness. It can be understood that the construction of the 3D local environment feature mesh provides standardized input for subsequent fusion.

[0073] In one embodiment of the present invention, see [reference] Figure 2 When performing multi-source mesh stitching and conflict resolution operations at the edge situation nodes, spatiotemporal stamps and UAV position and attitude information are added to each received 3D local environment feature mesh. Using the global coordinate system maintained by the edge situation nodes as a reference, all 3D local environment feature meshes are transformed to this global coordinate system. In the global coordinate system, overlap detection is performed on voxels describing the same spatial region in different 3D local environment feature meshes. For overlapping voxels, their comprehensive feature vectors are compared, and a confidence-weighted fusion algorithm is used to generate a new comprehensive feature vector. This confidence-weighted fusion algorithm dynamically calculates the confidence weight of voxel features in each mesh based on the accuracy of the acquisition sensor, the distance to the target during acquisition, and the data freshness. For voxels in non-overlapping regions, they are directly incorporated into the fusion result, ultimately forming the wide-area environment model covering a wider spatial range.

[0074] In practice, an edge situational awareness node deployed at a communication base station covers an urban intersection area. Three drones, designated UAV1, UAV2, and UAV3, fly in this area from the east, south, and west, respectively. The edge situational awareness node continuously receives 3D local environmental feature meshes from these three drones. In practice, each received 3D local environmental feature mesh is appended with a spatiotemporal stamp and drone position and attitude information. The spatiotemporal stamp includes the absolute time of data acquisition and the time offset of mesh generation. The drone position and attitude information comes from the latitude, longitude, altitude coordinates, pitch, roll, and yaw angles output by the airborne integrated navigation system. For example, a mesh of UAV1 is accompanied by a timestamp of "2026-03-27 10:00:00.125UTC" and attitude information of "N39.9°, E116.4°, 50m, pitch=2°, roll=1°, yaw=45°". In practice, the global coordinate system maintained by the edge situation nodes is used as the reference to transform all three-dimensional local environmental feature meshes into the global coordinate system. The origin of the global coordinate system is set at the phase center of the communication base station antenna, and the northeast-sky coordinate system is adopted. The transformation process is completed by the rotation matrix and translation vector calculated by the pose information. After the transformation, the voxel features from different UAVs have a unified spatial reference.

[0075] In the global coordinate system, overlap detection is performed on voxels describing the same spatial region in different 3D local environment feature meshes. Overlap detection is achieved by comparing whether the spatial indices of the voxels are consistent. In the example, the voxel with index (100, 200, 30) in the global coordinate system appears in both the meshes transmitted by UAV1 and UAV3, and these two voxels are determined to be overlapping. For overlapping voxels, their comprehensive feature vectors are compared, and a confidence-weighted fusion algorithm is used to generate a new comprehensive feature vector. In the example of the crossroads, for overlapping voxels describing the same building wall, UAV1, because its viewpoint is directly facing the wall, has significant texture contour features in its comprehensive feature vector, while UAV3, because its viewpoint is tilted and its distance is greater, has more prominent geometric planar features but blurred texture. The confidence-weighted fusion algorithm dynamically calculates the confidence weight of the voxel features in each mesh based on the acquisition sensor accuracy, the distance from the target during acquisition, and the data freshness of each 3D local environment feature mesh. In practical implementation, sensor accuracy is based on factory calibration parameters. LiDAR accuracy is higher than that of visible light cameras. The distance to the target during data acquisition is calculated from the UAV's position and voxel coordinates. Data freshness is determined by the difference between the current time and the spatiotemporal stamp. The process of generating a new comprehensive feature vector using the confidence-weighted fusion algorithm is expressed by the formula:

[0076]

[0077] Where: C fused It is the new integrated feature vector after fusion, Ci w is the original synthesized feature vector of the voxel in the i-th overlapping 3D local environment feature mesh. i is the confidence weight of the voxel feature in the i-th grid, and n is the number of overlapping grids. Confidence weight w i From sensor accuracy factor s i Distance factor d i and freshness factor t i The decision is made jointly, and its calculation method is w. i =α·s i +β·(1 / d i )+γ·(1 / (t now -t i +δ)), where s i It is the normalized sensor accuracy value, d i It is the Euclidean distance from the acquisition drone to its position in voxel space, t i It is the timestamp of data collection, t now α, β, and γ are the adjustment coefficients of each factor, and δ is a small constant added to prevent division by zero. In data comparison, for the same wall voxel, if simple averaging fusion is used, its texture and geometric features are weakened. However, after using the confidence-weighted fusion of this embodiment, because UAV1 data is updated and sensor accuracy is high, its weight w1 is higher than the weight w3 of UAV3, and the fusion result retains clearer and more reliable wall texture features. For voxels in non-overlapping areas, they are directly incorporated into the fusion result. For example, the bus stop area south of the intersection observed by UAV2 is not covered by the fields of view of UAV1 and UAV3. All voxels in this area are directly added to the fusion grid, ultimately forming a wide-area environment model covering a wider spatial range in the east, south, and west directions of the intersection. In some embodiments, the threshold for overlap detection can be set to half the voxel size to handle registration errors. In some embodiments, data consistency can be introduced as an additional factor in the confidence weight calculation. Optionally, the weight of the data freshness factor γ of the voxel where the dynamic obstacle is located can be increased. Optionally, a geocentric map coordinate system can be used as the global coordinate system to support large-scale operations. It can be understood that multi-source mesh stitching and conflict resolution construct a unified environmental representation. It can also be understood that the wide-area environment model provides the foundation for subsequent situational simulations.

[0078] In one embodiment of the present invention, when performing future situation prediction based on the wide-area environment model, a pre-set semantic understanding rule is first invoked to identify and classify the elements in the model, marking static terrain structures, dynamic obstacles, and potential risk areas. The specific process of element identification and classification involves defining a set of semantic categories, which at least include building surfaces, vegetation canopies, moving vehicles, pedestrians, and high-voltage cables. A corresponding feature matching template is set for each semantic category, describing a typical combination of texture features, geometric features, and spatial distribution features for the corresponding semantic category. Each voxel or voxel set in the wide-area environment model is traversed, and the similarity between its comprehensive feature vector and the feature matching template of each semantic category is calculated. The voxels or voxel sets are assigned to the semantic category with the highest similarity, and a corresponding category label is assigned to each voxel or voxel set. Subsequently, using the wide-area environment model and the annotation results, a virtual physics engine is driven to perform short-term future situation prediction, forecasting the state changes of various elements in the environment over a future period. Based on the results of the future situation prediction, the airspace passage cost map of the UAV swarm at the current and future times is calculated. By combining the airspace access cost map with the UAV's scheduled flight mission, real-time obstacle avoidance and path replanning instructions are generated for each UAV. These instructions are then sent to the corresponding UAVs to update the target waypoints and flight strategies in their flight control systems.

[0079] In practical implementation, the edge situation node has completed the construction of a wide-area environment model of the intersection area. The first step in predicting future situations based on this wide-area environment model is to call pre-set semantic understanding rules to identify and classify the elements in the model, thereby marking static terrain structures, dynamic obstacles, and potential risk areas. A set of semantic categories is defined, including building surfaces, vegetation canopies, moving vehicles, pedestrians, and high-voltage cables, and a corresponding feature matching template is set for each semantic category. The feature matching template describes the typical combination of texture features, geometric features, and spatial distribution features of the corresponding semantic category in the form of vectors. The texture feature dimension includes color histogram and edge density; the geometric feature dimension includes curvature value and flatness; and the spatial distribution feature dimension includes horizontal continuity and vertical distribution range. An exemplary feature matching template library is shown in Table 1.

[0080] Table 1: Semantic Category Feature Matching Template Table

[0081]

[0082] For each voxel or set of voxels in the wide-area environment model, the similarity between its comprehensive feature vector and the feature matching template of each semantic category is calculated. The similarity calculation adopts a metric method based on weighted Euclidean distance, and the formula is expressed as:

[0083]

[0084] Wherein: S k F is the similarity score between the comprehensive feature vector of the current voxel or voxel set and the k-th semantic category template. m It is the value of the composite feature vector in the m-th feature dimension, T k,m λ is the typical value of the feature matching template for the k-th semantic category on the m-th feature dimension. m is the weight coefficient of the m-th feature dimension, where M is the total number of feature dimensions. In the intersection example, a set of voxels located on the ground with high flatness and regular edges, which, along with the S of the "building surface" template... 建筑物 The value could be -0.1, while the S value of the "Moving Vehicle" template is different. 车辆 The value may be -5.6, with higher scores (smaller negative values) indicating greater similarity. Voxels or sets of voxels are assigned to the semantic categories with the highest similarity, and corresponding category labels are assigned to these voxels or sets. In data comparison, without semantic understanding rules, the model only contains geometric and texture features and cannot distinguish between stationary vehicles and low-rise buildings. After applying the classification method of this implementation, stationary vehicles are correctly identified and labeled as "moving vehicles," and low-rise buildings are labeled as "building surfaces," demonstrating the accuracy of the classification.

[0085] Using a wide-area environment model and annotation results of labeled static terrain structures, dynamic obstacles, and potential risk areas, a virtual physics engine is driven to perform short-term future situation projection, predicting the state changes of various elements in the environment over a future period. Based on the results of the future situation projection, the airspace access cost map of the UAV swarm at the current and future times is calculated. Combining the airspace access cost map with the UAV's scheduled flight mission, real-time obstacle avoidance and path replanning instructions applicable to each UAV are generated. For example, if a UAV's scheduled flight mission is to navigate from the east side of an intersection to the west side, its target point sequence and time constraints are parsed out. The generated real-time obstacle avoidance and path replanning instructions are issued to the corresponding UAVs, updating the target waypoints and flight strategies in their flight control systems. In some embodiments, semantic categories can be expanded to include categories such as water bodies and construction areas. In some embodiments, similarity calculation can use cosine similarity or Mahalanobis distance. Optionally, voxel regions with low classification confidence can be marked as "unknown" for manual verification. Optionally, feature matching templates can be trained on a specific scene dataset through machine learning. It is understandable that semantic understanding associates geometric features with physical meaning. It is also understandable that future situational projections provide input for dynamic programming.

[0086] See Figure 3The heatmap presents the matching strength relationship between different semantic elements and three feature dimensions. The heatmap uses five core semantic elements, namely building surface, vegetation canopy, moving vehicles, pedestrians and high-voltage cables, as rows and three feature dimensions, namely texture features, geometric features and spatial distribution features, as columns. The cell value represents the feature matching degree between the corresponding semantic category and feature dimension (the value range is 1-9, and the higher the value, the stronger the matching). The following professional conclusions can be drawn from the numerical distribution: Building surface: It shows a very high matching degree in both geometric features (9) and texture / spatial distribution features (8), which conforms to its template definition of "regular edges, high flatness, and vertical concentrated distribution", reflecting the strong typicality of this category in the three features. High-voltage cable: It shows a prominent matching degree in geometric features (8) and spatial distribution features (9), and a very low matching degree in texture features (1), which accurately corresponds to its physical form of "linear low curvature structure, slender horizontal distribution and no obvious texture". The feature dimension has significant distinguishability. Vegetation canopy: The matching degree of texture and geometric features is 2, only improving to 6 in spatial distribution features, reflecting that its template features of "complex irregular texture, high curvature, and patchy distribution" have better spatial recognition. Moving vehicles: The matching degree of texture features (6) is the highest, while the matching degree of geometric (5) and spatial distribution (4) features decreases in turn, which matches its category attributes of "regular metal outline, medium curvature, and isolated distribution on the ground". Pedestrians: The matching degree of all three types of features is at a low level (2-3), reflecting the weak distinguishability of its features of "small-scale complex outline, non-planar, and point distribution" in template quantization. The heat map provides an intuitive weight reference for the subsequent similarity calculation of voxel sets and semantic templates, and is the core visual support for the feature dimension weight allocation in semantic understanding rules.

[0087] In one embodiment of the present invention, the process of driving the virtual physics engine to perform short-term future situation projection includes: extracting elements classified as dynamic obstacles from the wide-area environment model and establishing simplified dynamic models for them; extracting static terrain structures from the wide-area environment model as collision boundaries and motion constraints in the virtual physics engine; setting initial values ​​of motion states based on current observations for the dynamic models of each dynamic obstacle and assigning them behavioral rules consistent with their semantic categories; simulating gravity and wind resistance physical effects in the virtual physics engine and loading the collision boundaries and motion constraints; running the virtual physics engine to calculate the predicted positions and motion states of all dynamic obstacles in multiple consecutive time slices in the future, forming a future situation sequence; when calculating the airspace access cost map, using the wide-area environment model as the spatial basis, dividing it into three-dimensional spatial cells; defining a cost calculation function that comprehensively considers the distance between the cell and the static terrain structure, the distance between the cell and the dynamic obstacle at the prediction time, and the signal occlusion degree of the cell; and inputting the predicted positions of dynamic obstacles at each time step in the future situation sequence into the cost calculation function at the corresponding time step. For each future time slice, the access cost of each three-dimensional spatial cell is calculated, generating a static and dynamic integrated access cost distribution map for the corresponding future time slice. The sequence of integrated access cost distribution maps arranged in chronological order is then integrated into a spatiotemporal four-dimensional spatial access cost map.

[0088] In practical implementation, edge situation nodes drive the virtual physics engine to perform short-term future situation inference based on the labeled wide-area environment model. Elements classified as moving vehicles and pedestrians are extracted from the wide-area environment model as dynamic obstacles, and simplified dynamic models are established for them. For example, for moving vehicles, a mass model is used to represent their position and velocity, and for pedestrians, a social force model is used to describe their motion trend. The parameters of the dynamic model include mass, maximum velocity, and acceleration. Elements classified as building surfaces and streetlights are extracted from the wide-area environment model as static terrain structures, which serve as collision boundaries and motion constraints in the virtual physics engine. In the example urban road intersection scene, building polygons and streetlight cylinders are loaded as impenetrable colliders. For each dynamic obstacle's dynamic model, initial values ​​for its motion state based on current observations are set, and behavioral rules conforming to its semantic category are assigned. For example, the extracted vehicle dynamic model, numbered Car1, has an initial position at the east entrance of the intersection, an initial speed of 8 meters per second traveling west, and its behavioral rule is set to "travel along the lane direction, and may decelerate or turn at the intersection according to traffic light rules." The extracted pedestrian dynamic model, numbered Ped1, has an initial position at the southwest corner of the sidewalk, an initial speed of 1.2 meters per second moving north, and its behavioral rule is set to "move along the sidewalk towards the target point, avoiding collisions with static obstacles and other pedestrians." In the virtual physics engine, gravity and wind resistance physical effects are simulated, and collision boundaries and motion constraints formed by buildings and streetlights extracted from the wide-area environment model are loaded. The gravitational acceleration is set to 9.8 m / s². 2 The drag coefficient is set based on current meteorological data. A virtual physics engine is run to calculate the predicted positions and motion states of all dynamic obstacles in multiple consecutive time slices in the future, forming a future situation sequence. The time slice interval is 0.1 seconds, and the prediction time domain length is 2 seconds, generating a sequence containing 20 time slices. See Table 2 for some examples of the future situation sequence.

[0089] Table 2: Predicted Location Sequence of Dynamic Obstacles in the Next 2 Seconds

[0090]

[0091] In data comparison, if path planning is based solely on the static snapshot of the current moment, the system will assume that Car1 will forever occupy the airspace where it was initially located. However, this implementation method, through future situation projection, can predict that Car1 will move from coordinates (100.0, 50.0) to (92.0, 50.0) in the next second, thereby identifying the area as a temporarily occupied but soon-to-be-released dynamic space.

[0092] Based on the results of future situation projections, the airspace passage cost map of the UAV swarm at the current and future times is calculated. Using a wide-area environment model as the spatial basis, it is divided into three-dimensional spatial cells with sides of 1 meter. A cost calculation function is defined, which comprehensively considers the distance between the cell and static terrain structures, the distance between the cell and dynamic obstacles at the predicted time, and the signal obstruction degree of the cell. The signal obstruction degree is determined by whether there are buildings or other obstructions between the cell and the communication base station. The predicted positions of dynamic obstacles at each time point in the future situation sequence are input into the corresponding cost calculation function. For each future time slice, the passage cost of each three-dimensional spatial cell is calculated, generating a combined static and dynamic passage cost distribution map for the corresponding future time slice. The cost calculation function is expressed as:

[0093] C total (x,y,z,τ)=w s ·C s (d static )+w d ·C d (d dynamic (τ))+w o ·C o (visibility)

[0094] Where: C total (x,y,z,τ) represents the total passage cost at the cell with spatial coordinates (x,y,z) at prediction time τ, C s Based on the distance d between the cell and the nearest static terrain structure static The calculated static terrain cost term increases with closer distances, C d Based on the distance d between the cell at time τ and the predicted position of the nearest dynamic obstacle. dynamic (τ) is the calculated dynamic obstacle cost term; the closer the obstacle, the higher the cost. C o This is a signal occlusion cost term calculated based on the signal visibility of the cell to the base station. The cost is highest when there is complete occlusion. s ,w d ,w o These are the weighting coefficients for the three costs. In the example scenario, a cell adjacent to a building has a weighting coefficient due to its d... static Small value with high C s The cost is that a cell that overlaps with the predicted Car1 position at time T+1.0 has its d dynamic A value of zero with extremely high C d The cost is that a cell completely obscured by a building has the highest C. oCost. The sequence of comprehensive passage cost distribution maps arranged chronologically is integrated into a four-dimensional spatiotemporal airspace passage cost map. This map adds a time dimension to the three-dimensional space, recording the passage risk of each spatial location at different future times. In some embodiments, the cell size can be adjusted to 0.5 meters or 2 meters according to the task accuracy requirements. In some embodiments, the dynamic obstacle cost item C... d The relative velocity direction of the obstacle can be further considered. Optionally, the signal blocking cost C... o It can be calculated using a ray tracing algorithm. Optional, the weighting coefficient w. s ,w d ,w o The settings can be dynamically adjusted according to the type of UAV mission. For example, for reconnaissance missions that prioritize signal quality, the settings can be improved. o Weighting. It's understandable that the future situation sequence forms the basis for calculating dynamic costs. It's also understandable that the four-dimensional cost map integrates spatiotemporal risk information.

[0095] See Figure 4In the edge computing-based UAV airspace passage cost assessment system, the passage costs of dynamic obstacle areas and safe areas exhibit a significantly differentiated evolution pattern over future time slices. Specifically, the cost of dynamic obstacle areas (dot curve) continuously climbs within the prediction time domain of 0.0 to 1.0 seconds, reaching a peak of 0.98 at 1.0 seconds, reflecting the rapid accumulation of passage risk due to the approach of dynamic obstacles in this area within the next second. Subsequently, it rapidly declines within the 1.0 to 2.0 second range, dropping to 0.22 at 2.0 seconds, indicating that the dynamic obstacle has passed through the area and the passage constraint has been released. In contrast, the cost of safe areas (triangular curve) remains within a narrow range of 0.20 to 0.22 throughout the entire 2-second prediction time domain, with minimal fluctuations, demonstrating that this area is less affected by static terrain and dynamic obstacles, and its passage stability is significantly better than that of dynamic obstacle areas. From the perspective of cost evolution logic, the initial rise and subsequent fall trend of the cost in dynamic obstacle areas is highly coupled with the dynamic obstacle movement sequence derived from the virtual physics engine—obstacles continuously approach the target area within 0.0 to 1.0 seconds, causing the distance between the cell and the dynamic obstacle to continuously shorten, amplifying the weight of the dynamic cost term; after 1.0 seconds, the obstacle gradually moves away, and the dynamic cost term decays accordingly, eventually approaching the cost of the safe area. The stability of the safe area cost stems from its spatial location being far from the static terrain structure and the movement path of the dynamic obstacle, while the signal obstruction degree is low. The static terrain cost term, the dynamic obstacle cost term, and the signal obstruction cost term all remain at low levels, and the weighted sum of the three forms a stable low cost baseline. The curve intuitively verifies the effectiveness of the four-dimensional airspace passage cost map based on future situation projection: by capturing the temporal evolution characteristics of dynamic obstacles, it is possible to accurately identify dynamic risk areas that are "temporarily occupied and about to be released," providing a refined spatiotemporal risk basis for real-time obstacle avoidance and path replanning of UAV swarms, avoiding misjudgments and passage efficiency losses caused by relying solely on static snapshot planning.

[0096] In one embodiment of the present invention, when generating real-time obstacle avoidance and path replanning instructions, the predetermined flight mission of each UAV is analyzed, and its target point sequence and time constraints are extracted. Starting from the UAV's current position and ending at the mission target point, a spatiotemporal trajectory from the current time to the target time is searched in the airspace access cost map. The optimization objective of the search is to minimize the sum of the access costs of all spatiotemporal cells traversed by the entire spatiotemporal trajectory, while simultaneously satisfying the UAV's dynamic constraints. The search process employs the spatiotemporal A* algorithm in the four-dimensional airspace access cost map, treating time as an additional one-dimensional extended cost. The optimal spatiotemporal trajectory obtained from the search is calculated into a series of time-ordered spatial coordinate points and velocity instructions, constituting the real-time obstacle avoidance and path replanning instructions.

[0097] In practical implementation, a UAV-α needs to perform a material transfer task in an urban intersection area. The edge situation node has generated an airspace access cost map covering this area, based on which real-time obstacle avoidance and path replanning instructions suitable for the UAV-α are generated. The UAV-α's scheduled flight mission is analyzed to extract its target point sequence and time constraints. The scheduled flight mission is described as "flying from point A (coordinates (0,0,50)) on the roof of the warehouse building to point B (coordinates (200,150,60)) on the roof of the distribution station building within 60 seconds, and hovering at point C (coordinates (100,50,30)) during the flight, with the hovering time no later than 25 seconds after takeoff". After analysis, the target point sequence [A,C,B] and the corresponding time constraints are obtained. Starting from the current position of the UAV-α and ending at the mission target point B, a spatiotemporal trajectory from the current time to the target time is searched in the airspace access cost map.

[0098] The optimization objective of the search is to minimize the sum of the travel costs across all spatiotemporal cells traversed by the entire spatiotemporal trajectory, while simultaneously satisfying the UAV's dynamic constraints, including maximum speed, maximum acceleration, and minimum turning radius. The search process employs the spatiotemporal A algorithm within a four-dimensional airspace travel cost map, treating time as an additional one-dimensional expansion cost. The spatiotemporal A algorithm expands from a four-dimensional node representing the UAV's current state, with each node containing spatial coordinates (x, y, z) and time t. In specific implementation, the cost of expanding from the current node to the next neighboring node is... move Composed of the cost integral of the spatiotemporal cells traversed and a time penalty term, the heuristic function estimates the remaining cost from the current node to the target node. The total travel cost J(Γ) of a spatiotemporal trajectory Γ is the sum of all extended costs experienced from its starting point to its ending point, calculated as follows:

[0099]

[0100] Where: Γ represents a spatiotemporal trajectory consisting of N discrete state points, p k =(x k ,y k ,z k ) represents the spatial coordinates of the k-th point on the trajectory, and t represents the spatial coordinates of the k-th point k It is the time of arrival at that point. It is the airspace access cost map at coordinate p k and time t k The per-unit-length transport value at a given point is a dimensionless scalar, Δs. k =‖p k+1 -p k ‖ is from point p k Move to point p k+1 The Euclidean distance, whose dimension is length, Δtk =t k+1 -t k This represents the time spent completing this movement, where η is the penalty coefficient per unit time, a dimensionless scalar used to balance time efficiency with spatial travel risk. Under this definition, The term η·Δt represents the cumulative spatial risk cost of a path, and the term η·Δt represents the time consumption cost. Both terms are dimensionless, and the total cost J(Γ) is also a dimensionless scalar. In the intersection example, if a path traverses a high-risk area (high... Even if the path is short (small Δs), its cumulative cost can be high; if a detour through a low-risk area is chosen, the path may become longer, but the cumulative cost may be lower. The spatiotemporal A* algorithm finds the trajectory with the minimum total cost by evaluating the J(Γ) values ​​of different spatiotemporal paths.

[0101] In its implementation, the algorithm searches in four-dimensional space, and when expanding nodes, it must satisfy dynamic constraints, such as the current node having a velocity of v. k Then the next reachable position p k+1 Must satisfy ||p k+1 -p k || / Δt k ≤v max Under the same conditions. In the data comparison, if only the cost map at the current time is used for a three-dimensional spatial A search, the resulting path will pass through a point at t=12 seconds where the cost per unit length is increased due to the prediction of a vehicle passing through. In an area with a value as high as 100, the path segment length Δs is 5 meters, therefore the cost contribution of this segment is 500; while the four-dimensional trajectory planned by the spatiotemporal A algorithm passes through this area before t=10 seconds, at which point... Although the acceleration in advance slightly increases Δt, the total cost J(Γ) of the entire trajectory is lower and safer. The optimal spatiotemporal trajectory obtained by the search is solved into a series of time-ordered spatial coordinate points and velocity commands, which constitute real-time obstacle avoidance and path replanning commands. For example, the command sequence includes "at t=8.2 seconds, arrive at (38,5,52), and set the speed to (4.5,1.2,0.3)". In some embodiments, the unit time penalty coefficient η can be dynamically set according to the urgency of the task. For urgent tasks, a higher η value can be set to prioritize the path with shorter time. In some embodiments, the heuristic function can be designed as a linear combination of the Euclidean distance to the target point and the estimated remaining time. Optionally, when the search space is too large, a hierarchical search strategy can be adopted, first performing coarse-grained spatiotemporal planning, and then performing local fine-grained correction. Optionally, the velocity command needs to be smoothed when solving to meet the execution requirements of the UAV flight control system. It can be understood that the spatiotemporal A* algorithm searches in a four-dimensional cost space. It can be understood that the solved commands directly drive the UAV flight.

[0102] See Figure 5 In the comparative experiment of total cost of UAV path planning, the performance difference between the spatiotemporal A* algorithm and the 3D A* algorithm intuitively demonstrates the advantages of four-dimensional spatiotemporal cost modeling. The graph shows the following overall trend: the cumulative passage cost of both algorithms increases monotonically with flight time, consistent with the basic law of path cumulative cost. The total path cost of the spatiotemporal A* algorithm (dotted curve) is consistently significantly lower than that of the 3D A* algorithm (boxed curve). Within a 60-second flight cycle, the former's final cost is approximately 270, while the latter's is approximately 700, representing a cost reduction of over 60% for the spatiotemporal A* algorithm. Key event analysis: The dashed line in the graph marks the high-risk moment t=12s. Because the 3D A* algorithm only models the static environment at the current moment, it is forced to traverse the high-risk area at this moment, resulting in a steep jump in the cost curve (a single cost contribution of approximately 500). In contrast, the spatiotemporal A* algorithm, through advance planning, completes passage through this area before t=10s, avoiding the high-risk moment, and the cost increase remains smooth. Performance Essence: The surge in cost of the 3D A* algorithm stems from its inability to utilize future situational information, forcing it to passively respond to the current environment. The Spatiotemporal A* algorithm, however, incorporates time as a fourth dimension of extended cost, searching for the optimal spatiotemporal trajectory within the airspace passage cost map. By proactively avoiding risks in the time dimension, it achieves global optimization of both spatial risk and temporal efficiency. Engineering Significance: This result validates the value of wide-area environment models and future situational projections based on edge computing. The Spatiotemporal A* algorithm effectively utilizes four-dimensional cost information, generating safer and lower-cost flight paths while satisfying UAV dynamic constraints. This provides reliable technical support for autonomous obstacle avoidance and mission execution by UAVs in complex urban environments.

[0103] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for UAV flight environment reconstruction and situational awareness based on edge computing, characterized in that, The method includes: The multi-source sensor array carried by the UAV synchronously collects environmental perception information of the flight airspace, including visible light images, depth point clouds and radio frequency signal characteristics. The environmental perception information is input into the airborne fusion computing unit, and the visible light image, depth point cloud and radio frequency signal features are spatiotemporally registered in the airborne fusion computing unit to generate a registered heterogeneous data block. Feature extraction and rasterization mapping are performed on the registered heterogeneous data blocks to construct a three-dimensional local environment feature mesh centered on the UAV, including: Texture feature sub-blocks, geometric structure sub-blocks, and signal strength sub-blocks are separated from the registered heterogeneous data blocks; Edge and corner detection is performed on the texture feature sub-blocks to extract contour features from the image; The surface normal and curvature of the geometric structure sub-blocks are estimated and the plane and edge features in the point cloud are extracted. The signal source direction is estimated for the signal strength sub-block, and the spatial distribution characteristics of the radio frequency signal are extracted. Define a fixed-size three-dimensional voxel space, and map the extracted contour features, plane features, edge features, and spatial distribution features to the corresponding voxels according to their spatial coordinates; A comprehensive feature vector is calculated for each voxel filled with features, and all voxels and their corresponding comprehensive feature vectors together constitute the three-dimensional local environment feature mesh. The three-dimensional local environment feature mesh is transmitted to the edge situation node deployed at the communication base station, where the three-dimensional local environment feature mesh from multiple UAVs is received. Within the edge situation node, multi-source mesh stitching and conflict resolution operations are performed to fuse multiple 3D local environment feature meshes from different perspectives and times into a unified wide-area environment model, including: Add a spatiotemporal stamp and UAV position and attitude information to each received 3D local environment feature mesh; Using the global coordinate system maintained by the edge situation nodes as a reference, all three-dimensional local environment feature meshes are transformed to the global coordinate system; In the global coordinate system, overlap detection is performed on voxels describing the same spatial region in different 3D local environment feature meshes; For overlapping voxels, their comprehensive feature vectors are compared, and a confidence-weighted fusion algorithm is used to generate new comprehensive feature vectors. For voxels in non-overlapping regions, they are directly incorporated into the fusion result, ultimately forming the wide-area environment model that covers a wider spatial range; Based on the wide-area environment model, future situation simulation is performed to generate UAV flight commands; The confidence-weighted fusion algorithm dynamically calculates the confidence weight of voxel features in each grid based on the acquisition sensor accuracy, the distance from the target during acquisition, and the data freshness of each 3D local environment feature grid.

2. The UAV flight environment reconstruction and situational awareness method based on edge computing as described in claim 1, characterized in that, Based on the aforementioned wide-area environment model, future situation simulations are performed to generate UAV flight commands, including: Based on the wide-area environment model, preset semantic understanding rules are invoked to identify and classify the elements in the model, and static terrain structures, dynamic obstacles and potential risk areas are marked. Using the wide-area environment model and annotation results, a virtual physics engine is driven to perform short-time domain future situation simulation and predict the state changes of various elements in the environment over a period of time. Based on the results of the future situation projection, calculate the airspace passage cost map of the drone swarm at the present and future times; By combining the airspace access cost map with the UAV's scheduled flight mission, real-time obstacle avoidance and path replanning instructions are generated for each UAV. The real-time obstacle avoidance and path replanning commands are sent to the corresponding UAVs to update the target waypoints and flight strategies in their flight control systems.

3. The UAV flight environment reconstruction and situational awareness method based on edge computing as described in claim 2, characterized in that, Based on the aforementioned wide-area environment model, pre-defined semantic understanding rules are invoked to identify and classify the elements in the model, specifically including: Define a set of semantic categories, which at least include building surfaces, vegetation canopies, moving vehicles, pedestrians, and high-voltage cables; A corresponding feature matching template is set for each semantic category. The feature matching template describes the typical combination of texture features, geometric features and spatial distribution features of the corresponding semantic category. Traverse each voxel or set of voxels in the wide-area environment model and calculate the similarity between its comprehensive feature vector and the feature matching template of each semantic category. Voxels or sets of voxels are classified into the semantic categories with the highest similarity, and the corresponding category labels are assigned to the voxels or sets of voxels.

4. The UAV flight environment reconstruction and situational awareness method based on edge computing as described in claim 3, characterized in that, Using the aforementioned wide-area environment model and annotation results, a virtual physics engine is driven to perform short-time domain future situation prediction, including: Elements classified as dynamic obstacles are extracted from the wide-area environment model, and simplified dynamic models are established for them; Static terrain structures are extracted from the wide-area environment model and used as collision boundaries and motion constraints in the virtual physics engine. For each dynamic obstacle, set initial values ​​for its motion state based on the current observations and assign it behavioral rules that conform to its semantic category. In the virtual physics engine, the physical effects of gravity and wind resistance are simulated, and the collision boundaries and motion constraints are loaded. The virtual physics engine is run to calculate the predicted positions and motion states of all dynamic obstacles in multiple consecutive time slices in the future, forming a future situation sequence.

5. The UAV flight environment reconstruction and situational awareness method based on edge computing as described in claim 4, characterized in that, Based on the results of the future situation projection, the airspace passage cost map of the drone swarm at the current and future times is calculated, including: Using the aforementioned wide-area environment model as a spatial basis, it is divided into three-dimensional spatial cells; Define a cost calculation function that comprehensively considers the distance between the cell and the static terrain structure, the distance between the cell and the dynamic obstacle at the prediction time, and the signal occlusion degree of the cell. The predicted location of the dynamic obstacle at each moment in the future situation sequence is input into the cost calculation function at the corresponding moment. For each future time slice, calculate the passage cost of each three-dimensional spatial cell and generate a static and dynamic integrated passage cost distribution map for the corresponding future time slice; The sequence of comprehensive traffic cost distribution maps arranged in chronological order is integrated into the aforementioned four-dimensional spatiotemporal airspace traffic cost map.

6. The UAV flight environment reconstruction and situational awareness method based on edge computing as described in claim 5, characterized in that, Combining the airspace access cost map with the UAV's scheduled flight mission, real-time obstacle avoidance and path replanning instructions are generated for each UAV, including: Analyze the planned flight mission of each drone and extract its target point sequence and time constraints; Starting from the current location of the UAV and ending at the mission target point, search for a spatiotemporal trajectory from the current time to the target time in the airspace passage cost map. The optimization objective of the search is to minimize the total travel cost of all spatiotemporal cells traversed by the entire spatiotemporal trajectory, while simultaneously satisfying the dynamic constraints of the UAV. The optimal spatiotemporal trajectory obtained from the search is solved into a series of spatial coordinate points and velocity commands ordered by time, which constitute the real-time obstacle avoidance and path replanning commands.

7. The UAV flight environment reconstruction and situational awareness method based on edge computing as described in claim 6, characterized in that, The search process employs the spatiotemporal A* algorithm in a four-dimensional spatial travel cost map, with time used as an additional one-dimensional extended cost.

Images