Drone Three-Dimensional Low-Altitude Urban Operation Guidance Method Based on Lightweight Task Templates

Abstract

The present invention discloses a method for guiding the low-altitude urban operation of an unmanned aerial vehicle based on a lightweight task template, including: after the UAV-Agent enters the geographical fence, it collects environmental information and constructs a three-dimensional rasterized graph structure, where homogeneous nodes represent key points of the planar path, and heterogeneous nodes represent the three-dimensional space path relationship; dynamically updates the task template based on mobile edge computing and stores the guiding path information; the user machine requests a path from the navigation station through lightweight interaction, the navigation station matches the optimal guiding path and feeds it back, and the user machine autonomously navigates along the key points. The innovation of the present invention lies in: adopting a heterogeneous node graph structure to model the three-dimensional airspace of the city to achieve multi-layer path planning; based on the dynamic update mechanism of the lightweight task template, reducing communication overhead; combining the key point optimization algorithm of geometric topology to improve navigation safety. Compared with the traditional route planning, this solution can significantly reduce the consumption of computing and communication resources.

Description

Technical Field

[0001] The present invention relates to drone control, and specifically to a method for guiding three-dimensional low-altitude urban operation of drones based on a lightweight task template. Background Art

[0002] In existing drone control solutions, the flight path planning of drones usually requires static map data to obtain the geographical basic information of each flight point in the initial path, such as longitude, latitude, and altitude information. However, in complex urban environments, drones used for logistics, express delivery, inspection, and area coverage require dynamic drone control and dynamic path planning solutions. Therefore, the drone control system needs to perform periodic information interaction with multiple map data APIs through a mobile edge server to obtain geographical basic information different from that of the flight points on the initial path. For ground unmanned vehicles, this information interaction may solve existing problems in path planning. However, for drones in complex three-dimensional urban environments, this information interaction requires a large amount of communication and computing resources; furthermore, precise drone planning requires the creation and access of a larger-scale high-precision dynamic map platform. The high-precision dynamic map platform is considered the core support for the safe and efficient operation of drone logistics and express delivery. Currently, there is no completely mature high-precision dynamic map platform dedicated to drones.

[0003] To meet the requirements of logistics, express delivery, inspection, and area coverage in complex urban environments, the flight path planning of drones requires static and dynamic time-space data. However, existing high-precision dynamic map platforms and bird's-eye view (BEV) databases are not yet mature, which limits the efficiency and safety of low-altitude urban planning for drones. Summary of the Invention

[0004] Object of the Invention: Aiming at the above disadvantages, the present invention provides a method for guiding three-dimensional low-altitude urban operation of drones based on a lightweight task template with high control efficiency.

[0005] Technical Solution: To solve the above problems, the present invention adopts a method for guiding three-dimensional low-altitude urban operation of drones based on a lightweight task template, including the following steps:

[0006] (1) The drone intelligent agent UAV-Agent enters the geofence and sends a starting point collection instruction to the navigation station within the geofence;

[0007] (2) After receiving the starting point collection instruction, the navigation station feeds back the starting point and ending point information within the geofence and the correct reception information ACK to the UAV-Agent;

[0008] (3) The UAV-Agent plans an initial path based on the feedback information and runs from the starting point to the ending point. During the running process, it collects the environmental information within the geographical fence, constructs a preliminary task template according to the collected environmental information, and updates the task template through the mobile edge server.

[0009] The task template includes a graph structure obtained by rasterizing a three-dimensional low-altitude area, several path key points, and several guiding paths planned according to the graph structure and path key points.

[0010] (4) Send the updated task template to the navigation station. The navigation station uses a path planning algorithm for the path key points on the guiding paths in the task template to plan the guiding paths and obtain a smooth navigation path.

[0011] (5) The user machine at the starting point in the geographical fence sends a task request to the navigation station.

[0012] (6) After receiving the task request, the navigation station selects the guiding path data from the task template and feeds back the navigation path to the user machine.

[0013] (7) After receiving the guiding path data and the navigation path, the user machine starts the subsequent process of autonomous navigation and sequentially passes through the path key points on the guiding path.

[0014] (8) When the current navigation task ends, the user machine notifies the navigation station and the navigation station feeds back ACK to the user machine.

[0015] Further, the task template includes the task type and corresponding process, the longitude, latitude, and altitude of the navigation station, the geographical fence boundary, the longitude, latitude, and altitude of the starting point and the ending point within the geographical fence, the number of guiding paths, the longitude, latitude, and altitude of several path key points, several guiding paths, the set of area IDs after rasterizing the three-dimensional low-altitude area, the area ID index, and the area topological relationship.

[0016] Further, the construction steps of the graph structure in the task template are as follows:

[0017] (31) Divide the three-dimensional low-altitude urban structure within the geographical fence into multiple grids with geometric topological relationships.

[0018] (32) Map the grids into multiple virtual roads, abstract a virtual road as a graph structure of isomorphic nodes, and store a single path, the path key points on the path, and the graph calculations related to path planning in each graph structure of isomorphic nodes; the graph structure describes the connection relationships between different regions in the environment; the virtual roads at different heights are abstracted as graph structures of heterogeneous nodes, and the graph structures of heterogeneous nodes include several paths, and the height of each path is different.

[0019] (33) Assign weights to the edges of the graph structure.

[0020] Further, the path planning algorithm includes an improved RRT-star algorithm; the improved RRT-star algorithm assists in sampling the sampling points in the RRT-star algorithm through path key points, specifically:

[0021] Successively use the path key points on the guiding path as prior positions;

[0022] For the interval between the starting point and the path key point, obtain the prior position, find the node closest to the prior position, confirm the coordinates of the target point, take the rectangular area formed by the closest node and the target point as the sampling area, and perform target-biased sampling within the sampling area to sample a new node;

[0023] For the interval between two path key points, calculate a new local target position through the scaling factor, prior position, and coordinates of the closest node, find the node closest to the new target position, take the rectangular area formed by the node closest to the new target position and the local target position as the sampling area, and perform target-biased sampling within the sampling area to sample a new node.

[0024] Further, calculate the synthesis cost of the trajectory planned by the improved RRT-star algorithm, and update the weights of the edges in the graph structure through the synthesis cost; the synthesis cost The calculation formula is:

[0025] ;

[0026] Wherein, is the height value of node , is the height value of node , is the energy consumption function caused by the height change, is the distance between the trajectory and the nearest obstacle, , and are weight coefficients.

[0027] Further, the path planning methods adopted by the initial path in step (3) include: A* algorithm, Dijkstra algorithm, RRT algorithm; the environmental information collected during the operation includes image acquisition, location information request, and geographical API call to obtain data of path key points;

[0028] After the UAV-Agent reaches the end point, it generates an initial task template from the collected information and uploads it to the database of the Mobile Edge Computing (MEC) server at the same time. The MEC server calibrates and updates the initial task template and then feeds back the updated task template to the UAV-Agent. The UAV-Agent downloads the updated task template to the navigation station. After the UAV-Agent receives the ACK sent by the navigation station, it leaves the service area involved in the geographical fence.

[0029] Furthermore, the task template further includes spatial limit boundaries, traffic data, and meteorological data.

[0030] Furthermore, in step (6), after receiving the task request, when the navigation station collects real-time congestion data, the navigation station will perform graph calculation to obtain a sub-optimal guiding path.

[0031] The present invention also adopts a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the above method are implemented.

[0032] The present invention also adopts a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above method are implemented.

[0033] Advantageous effects: Compared with the prior art, the present invention has the remarkable advantage that: through the navigation station containing a lightweight task template, it effectively supports low-altitude UAV navigation and path planning. The user machine only needs to interact with the navigation station once, instead of periodically and dynamically receiving instructions from the dynamic map API. This helps to save communication and computing resources and improve the efficiency of UAV control. By using prior key points for path planning, the safety of low-altitude UAV urban planning is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a schematic flowchart of information collection and processing in the lightweight mechanism of the present invention.

[0035] Figure 2 is a schematic diagram of the information transfer process in the navigation task of the present invention.

[0036] Figure 3 is a schematic diagram of the flight path of a low-altitude UAV or the driving path of a ground vehicle in the present invention.

[0037] Figure 4 is a schematic diagram of the simulation process of the graph structure in the present invention.

[0038] Figure 5 is a schematic diagram of the graph structure of heterogeneous nodes in the present invention.

[0039] Figure 6 It is a schematic flow chart of the single key point-guided sampling method 1 in the present invention.

[0040] Figure 7 It is a schematic flow chart of the double key point-guided sampling method 2 in the present invention. Detailed implementation manners

[0041] In this embodiment, a three-dimensional low-altitude urban operation guidance method for drones based on a lightweight task template includes the following steps:

[0042] As Figure 1 shown, the process of lightweight geographic information acquisition and processing is carried out. This process is not used to construct a high-precision dynamic map, but to construct a low-altitude drone navigation knowledge base for applications such as logistics, express delivery, and inspection. Figure 1 The information flow of each entity in

[0043] (1) After the UAV-Agent entering the geofence sends a starting point acquisition instruction to the navigation station, the navigation station feeds back the correct reception information (ACK) to the UAV-Agent. Then, the UAV-Agent will fly from the starting point to the ending point along an initial path, which can come from path planning methods such as the A* algorithm, Dijkstra algorithm, RRTs algorithm, etc. At the same time, the UAV-Agent with conventional image processing or specific embodied intelligence will perform image acquisition, position information request, and geographic API call, and assign the key point data in the path to the BEV image attributes to complete the attribute annotation of the key points in the graph structure. When ID2 takes 0, it means the UAV-Agent turns on the camera to perform BEV image acquisition. The map API call accesses the remote geographic database through the MEC server.

[0044] (2) After the UAV-Agent reaches the ending point, it ends the image acquisition and generates an initial task template, which is simultaneously uploaded to the database of the MECserver. Optionally, if the UAV-Agent has specific embodied intelligence, it can real-time complete the visual language navigation function, such as using computer vision to achieve environmental perception and target recognition, parsing the navigation intention in the text, and uploading an updated task template. When ID2 takes 3, its navigation intention summary is: determine the first path key point, and the first path key point is within the line of sight of the starting point. Optionally, the visual language navigation function can convert voice instructions into text to facilitate generating path key points under other conditions by combining visual and language information.

[0045] The task template is a subclass in the standardized interface, which defines the standardized interface to facilitate ensuring interoperability between modules and reducing the integration difficulty. The task template facilitates the interoperability of each module in the large language model (LLMs) and Agent-based UAV control system. The UAV control system includes: a user machine, a navigation station, a UAV intelligent agent (UAV-Agent), and a mobile edge server (MEC Server).

[0046] The user machine is used to represent autonomous navigation vehicles on the ground and low-altitude UAVs (Unmanned Aerial Vehicles), and they can communicate with the navigation station through a wireless interface. The Navigation Station integrating wireless and wired interfaces can communicate with the UAV-Agent and can also communicate with the MEC Server. Table 1 shows a semantic entity, namely the task template, which is involved in the semantic understanding, data collection, and processing process of the navigation scenario.

[0047] Table 1 Initial Task Template

[0048]

[0049] The set of IDs for the collection and navigation tasks of ID1 is {0, 1, …, M_ID1}. When ID1 takes 0 and 1, they represent the collection task and the navigation task respectively.

[0050] The set of IDs for the navigation and operation tasks of ID2 is {0, 1, …, M_ID2}. When ID2 takes 0 and 1, they represent Process 1 adopted during the collection task (see Figure 1 ) and Process 2 adopted during the navigation task (see Figure 2 ). When ID2 takes 1, the UAV-Agent turns on the camera for environmental perception to obtain BEV and calculates the building height. Optionally, the UAV-Agent turns on the lidar sensor to more accurately estimate the building contour and height value. When ID2 takes 2, based on Process 2, the UAV in the navigation task turns on the camera for environmental perception, data processing, and dynamic autonomous navigation.

[0051] In this embodiment, the case when ID2 takes 3 is discussed in more detail because it requires the UAV-Agent to have a specific visual language navigation function, which depends on the type of LLMs dense model it uses. When ID2 takes 3, based on Process 2, in the geofence, the first path key point is on the starting line of sight (see Figure 3 ). The case when ID2 takes 4 is optional. The navigation station in the navigation task interacts with the UAV-Agent, and it can dynamically update the knowledge base in the navigation station to handle sudden low-altitude traffic congestion events.

[0052] ID3 contains the longitude, latitude, and altitude of the navigation station, the longitude, latitude, and altitude of the starting point and the ending point, the geographical fence boundary and the set of region IDs, and the number of guiding paths. Here, the region is an extension of the traditional three-dimensional grid concept. See Figure 4 , the projections of multiple successive three-dimensional grids on a guiding path are multiple rectangular regions and have the same height parameter; see Figure 5 , the three-dimensional grids with the same height parameter are defined as regions. The number of guiding paths within the geographical fence is M_ID3, the longitude, latitude, and altitude of the navigation station are the tuple (x, y, z), and the guiding path represents the path from the starting point to the ending point.

[0053] ID4 contains the guiding path i, the region ID index, and the region topological relationship.

[0054] ID5 contains the number of key points in the guiding path i.

[0055] ID6 contains multiple types of key points, such as the longitude, latitude, and altitude of path key points, low-altitude obstacles (such as wires, trees), and ground building type obstacles, etc. On a guiding path, these key points are in sequence.

[0056] ID7 contains the boundary of airspace restrictions (such as no-fly zones), traffic data, and corresponding meteorological data. Here, it should be noted that the boundary of airspace restrictions (such as no-fly zones) and the geographical fence boundary can be conceptually interchanged. When the geographical fence boundary is physically existent, it is the airspace restriction boundary; when the geographical fence boundary is virtual, a boundary is added in the BEV map for region division, annotation, and the definition of the search interval of the subsequent RRT-star.

[0057] Optionally, the definitions of ID8 and those after it are similar to the definitions of ID4 - ID7. However, the guiding path i + 1 of ID8 may correspond to other path planning requirements, such as the UAV path planning at other altitudes, the UAV path planning for low-altitude emergency roads, etc. LLMs and agents can be used to help understand different path planning requirements because the inputs of these requirements may be text and images. Thus, the geographical fence boundary definition of ID3 in Table 1 can adopt a hybrid linked list. This linked list can contain the annotations of each region and the key points in each region; it can also contain a link to associate with a BEV map with semantic annotations.

[0058] The UAV-Agent is responsible for basic low-altitude geographic data collection and partial data processing. When the number of guiding paths is greater than 1, the guiding path i and the guiding path i+1 will adopt different path trajectories, which makes the linked list length in Table 1 continuously increase and also leads to a large storage and search complexity for guiding paths and key points. To facilitate data processing and analysis in geographic information collection, the UAV-Agent needs to adopt LLMs and Agents to effectively extract geographic information parameters from the linked list and facilitate interoperability between modules in the involved UAV navigation system.

[0059] Table 1 divides the low altitude into multiple grids with geometric topological relationships. Each cuboid grid is represented by the longitude and latitude of the lower left vertex and the upper right vertex and the height of its lower left vertex. For the sake of evaluation, it is assumed that the vertical range d_h of each grid is the same. Thus, d_h is the distance between two adjacent neighbors in the vertical direction or the lane safety interval. This forms virtual lanes in the urban low altitude.

[0060] The grid is mapped into a graph structure. Table 1 mainly defines the semantic information of low-altitude urban planning, and the graph structure is a data structure that facilitates the storage, search, and calculation of geographic information data for low-altitude urban planning. As Figure 4 shown, from top to bottom, it successively represents the grid, the graph structure, and the simulation of finding guiding paths or trajectories in low-altitude urban planning.

[0061] With the assistance of key points stored in the graph structure, the RRT-star algorithm quickly finds the shortest path in terms of direction; the output of the RRT-star algorithm will update the edge weights in the graph structure; then graph calculations are performed to recalculate the combined cost of path planning and the combined cost of evaluating the switching of UAVs at adjacent altitudes. Further, since the LLM agent at the navigation station can process the collected measured data, the static graph structure in the navigation knowledge base is updated and graph calculations are performed in sequence, and the output of the latter will include the shortest path and the switching prediction of the adjacent low-altitude roads, etc.

[0062] In short, the above expressions form a UAV navigation method proposed in this embodiment, which is different from the previous ones, that is, a process that allows the navigation system to perform real-time simulation, monitoring, and performance optimization, and its closed-loop advantage enables the static graph structure to adapt to dynamic spatio-temporal traffic.

[0063] The global access order defines the geometric relationships of four adjacent grids. For altitude i, the rectangular area order from left to right is area 0, area 1, area 2, and area 3. The left edge lines of area 1 and area 2 are the same, which indicates that the right adjacent area of area 1 and area 2 is area 0, and the semantic meaning on the map is that there is an intersection at the position of the left edge line of area 0, and the left and right branches of (area 0) reach area 1 and area 2 respectively. The roads at altitude i and altitude i+1 have an up-down topological relationship.

[0064] Implementation of isomorphic graphs: In a static graph structure of isomorphic types, each node represents a rectangular area (a projection of a three-dimensional grid) with a lane safety interval. For node attributes, refer to the following constructor __init__.

[0065] def __init__(self, rect_id, attr0=Node((0,0)), attr1=Node((0,0)),attr2=0, attr3=0, attr4=0, attr5=Node((0,0)), attr6=Node((0,0)), attr7=Node((0,0))):

[0066] The constructor __init__ is used to initialize an instance of the Node_rect class, and its definition can be further optimized according to actual needs. It accepts parameters such as:

[0067] rect_id: The unique identifier of the rectangular area.

[0068] attr0: The longitude and latitude of the lower left corner of the rectangular area, with a default value of (0,0).

[0069] attr1: The longitude and latitude of the upper right corner of the rectangular area, with a default value of (0,0).

[0070] attr2: The number of reference points within the rectangular area, with a default value of 0.

[0071] attr3: The height of the rectangular area, with a default value of 0.

[0072] attr4: The congestion index of the rectangular area, such as the number of obstacles, with a default value of 0.

[0073] attr5: The longitude and latitude of reference point 0, with a default value of (0,0).

[0074] attr6: The longitude and latitude of reference point 1, with a default value of (0,0).

[0075] attr7: The longitude and latitude of reference point 2, with a default value of (0,0).

[0076] The definition of edge types in a dynamic graph structure usually includes factors such as connection relationships (such as the connection between roads and intersections), traffic flow, speed limits, and congestion levels. In the evaluation experiment, synthetic cost is used to enable the definition of a static isomorphic graph to be extended to a dynamic heterogeneous graph.

[0077] Implementation of Heterogeneous Graph: By a method similar to that of homogeneous graph, a directed acyclic graph (DAG) is constructed to simulate the relationship of rectangular regions (with lane safety intervals) in a three-dimensional geographical space. See Figure 5 , an example implementation of a heterogeneous graph for evaluating road switching is outlined below, which includes defining node classes and rectangular node classes, creating nodes and constructing connection relationships, creating a graph structure using NetworkX, merging graph structures, plotting graphs using Matplotlib, and calculating the shortest path, etc.

[0078] The functional description of creating nodes and constructing connection relationships is as follows: Two sets of nodes, node0_0, node1_0, node2_0, node3_0 and node0_1, node1_1, node2_1, node3_1, are created, representing rectangular regions on two height levels respectively.

[0079] The functional description of creating a graph structure using NetworkX is as follows:

[0080] Two directed graphs, G0 and G1, are created, representing the graph structures on two height levels respectively. Nodes and edges are added to the graphs, and weights are assigned to the edges.

[0081] The functional description of merging graph structures is as follows:

[0082] A new directed graph G is created, and the nodes and edges of G0 and G1 are added to G. Edges between G0 and G1 are added to represent the switching of the drone between different height levels.

[0083] The functional description of plotting graphs using Matplotlib is as follows:

[0084] The positions pos and labels of the nodes are defined. Different shapes and colors are used to represent ordinary nodes and heterogeneous nodes. Nodes and edges are plotted, and the labels at the centers of the nodes are displayed.

[0085] The functional description of calculating the shortest path is as follows:

[0086] The shortest paths and path lengths from node 0 to node 4, from node 4 to node 7, and from node 7 to node 3 are calculated using a path planning algorithm. These paths are merged to obtain the complete path and total length from node 0 to node 3 passing through node 4 and node 7.

[0087] Optionally, path planning based on heterogeneous graphs can consider various factors, such as: shortest path, least time, least congestion, etc. Note that, in the actual path planning process of drones greater than 10 meters, the local environmental information detected by cameras and lidar is usually fused to further improve the performance of the path planning scheme based on a single feature (Euclidean distance).

[0088] Heterogeneous graph-based path planning can effectively reduce the dependence on predicted trajectories based on prior knowledge, i.e., one or more key points on the guiding path, while the usual predicted trajectories contain a large number of sampling points. If the time-varying nature of the low-altitude obstacle positions cannot be ignored, the design of the UAV navigation system should also consider how to handle the problem of invalid reference points in the predicted trajectories.

[0089] Specifically, the output of the prior RRT-star simulation shown in this embodiment will be fed back into the graph structure, enabling the edge weights to be updated according to the measured data.

[0090] Considering that the accuracy of BEV scene understanding based on conventional image processing is relatively poor, this embodiment proposes a key point selection strategy. Considering that BEV scene understanding can roughly identify the four edge lines such as the top, bottom, left, and right of a rectangular area, selecting a position with a horizontal coordinate slightly greater than the midpoint of the left edge line of the rectangular area as the first key point for each rectangular area can, to a certain extent, avoid the recognition errors of BEV scene understanding. Of course, more reference points can be taken for each rectangular area. Figure 3 The horizontal range of area 0 is greater than 150 meters, so area 0 contains 3 reference points (starting point 1, path key point 9, navigation station 10). Area 0 contains building type obstacles (3, 4, 5, 6), low-altitude UAV suspension type obstacles 7, and low-altitude wire type obstacles 8. In addition, LLMs and Agent can be used to parse the navigation intent in the text and improve the accuracy of BEV scene understanding.

[0091] In Figure 3 the prior RRT-star algorithm can find the trajectory from starting point 1 to end point 2, and its output composite cost (see the subsequent definition) is fed back into the graph structure. Assuming that the UAV altitude is constant at h1, the three-dimensional grid is projected into a two-dimensional rectangular area. The rectangular areas passed by the trajectory are area 0 (region 0), area 1, and area 3 respectively, because it is assumed that the congestion level in area 2 is high. In a future urban ultra-low-altitude scenario, such as in the height range from 1 meter to 45 meters, in addition to static obstacles, the UAV flight will also encounter a large number of unpredictable dynamic obstacles, which challenges the timeliness of the output trajectory of the static path planning algorithm and also makes it difficult to achieve the path planning goal accurate to the sub-meter level and sub-second level. Therefore, according to the actual requirements, the actual path planning algorithm needs to run regularly to obtain the latest navigation trajectory.

[0092] To quickly give a usage demonstration of a task template (only when ID2 takes 3), in the geofence, the first path key point is on the line of sight of the starting point. The first path key point and the second path key point are respectively called UAV prior position 1 and prior position 2, and the navigation station is located at prior position 2.

[0093] Compared with the benchmark RRT-star (see RRT-star and its improvements below), the example modifies steps (2) to (4) of the benchmark RRT-star, reducing the sampled random points by using the key points or prior positions from the guiding path data. Secondly, it also integrates goal-biased sampling. It is a key-point-guided RRT-star path planning method, and its detailed design is as follows.

[0094] Considering the scenario of low-altitude UAV autonomous navigation in urban streets or urban canyons, the UAV faces challenges such as the lack of high-precision dynamic maps, weak GPS signals and cellular communication signals, and changing spatio-temporal low-altitude interferences different from ground navigation. A UAV path planning framework combining a graph structure and the RRT-star algorithm is considered. This framework not only provides key points on the sub-optimal path for the UAV, but also can generate a temporary path in time when the pre-planned path cannot be used. Even at the starting point only, the user's machine needs to interact with the navigation station once, instead of receiving instructions from the dynamic map API periodically and dynamically, which helps to save communication and computing resources. Using prior key points, both the UAV and the navigation station can perform path planning simulations to calculate the shortest path before navigation. However, since the navigation station can obtain the real-time road conditions beyond the line of sight, the simulation at the navigation station can obtain more effective trajectory predictions, thus playing the role of traffic lights in urban low altitudes.

[0095] The steps of the benchmark RRT-star (Rapidly-Exploring Random Tree Star algorithm) include: (1) Initialize the tree, that is, take the starting point as the root node of the tree; (2) Sample a random point, that is, randomly generate a point in the search space; (3) Expand the tree, that is, find the nearest node in the tree and expand a new node towards the random point; (4) Reconnect, that is, check the nodes within a certain range around the new node and try to reconnect through the new node to reduce the path length; (5) Update the path cost, that is, update the path costs of all affected nodes in the tree. Steps (2) to (4) are repeatedly iterated until an optimal path from the starting point to the target is found as the navigation path.

[0096] The advantages of the benchmark RRT-star based on Euclidean distance, probabilistic sampling, and incremental sampling are that it can quickly find an initial path and then continuously optimize it as the number of samples increases. Their computational complexity is moderate, and they can find solutions without using explicit information about obstacles in the configuration space. They rely on a collision checking module and construct a roadmap of a feasible trajectory by connecting a set of points sampled from the obstacle-free space.

[0097] Generally speaking, in terms of obstacle avoidance requirements, flight dynamics constraints, environmental complexity, and real-time performance, etc., the benchmark RRT-star still needs to be extended to effectively handle the safety requirements and flight constraints of drones in complex urban three-dimensional environments. In particular, the idea of visual language navigation is introduced into the information flow and device of the drone navigation system, and an example of RRT-star path planning assisted by key points is proposed. Its innovative strategy lies in the following single-key-point-assisted sampling method 1 and double-key-point-assisted sampling method 2. See Figure 3 , at the beginning of the navigation task, after the navigation station selects one or more guiding path data according to the requirements, it is sent to the user machine. See the case where ID2 takes 3 in Table 1. The guiding path data contains the definition of key points. The sampling method 1 is used from the starting point to the first key point, and the sampling method 2 is used from the first key point to the second key point. When the number of key points is greater than 2, method 2 is used in a loop.

[0098] For the navigation task, a key-point-guided sampling method 1 is proposed to provide a more direct straight-line path for the drone with a certain probability. Using the prior provided by a single key point (SingleKeypoint), this local node expansion method can quickly discover the initial path to reduce random space exploration and avoid redundancy in local sampling of the benchmark RRT-star. This strategy usually needs to combine the outputs of cameras or lidar to quickly discover routes longer than 10 meters.

[0099] The prior position is used to guide the sampling process, which is different from the step of sampling random points in the benchmark RRT-star algorithm. By sampling near the prior position, the user machine can more effectively explore the area of interest and improve the efficiency of path planning or graphics processing. When ID2 takes 3, the prior position 1 is the key point 1, and the prior position 2 is the key point 2 (navigation station position). When ID2 takes 4, it is a functional extension of the case where ID2 takes 3, which will not be discussed in detail here. The functions of the single-key-point-guided sampling method 1 and the double-key-point-guided sampling method 2 are described in turn below, and a sampling strategy for crossing the boundary between two regions is also mentioned at the end of the specification.

[0100] Function parameters of the single-key-point-guided sampling method 1:

[0101] There is only one parameter self in the definition of the method. It is used in this function to access and operate on the instance variables and methods of the RrtStar class.

[0102] Method implementation:

[0103] 1) Method definition:

[0104] This is an instance method. It has no parameters (except the above-mentioned self) and will operate using the attributes of the class.

[0105] 2) Calculate the prior index

[0106] self.prior_locations is a list containing prior locations, and self.index_ref is a reference index (in the implementation of this patent, it refers to the key point index). By calculating len(self.prior_locations) - self.index_ref through subtraction, an index of a prior location is obtained. When self.index_ref < 0, it indicates that all key points have been used up, and the proposed key point-guided RRT-star will execute steps (2), (3), (4), and (5) of the benchmark RRT-star.

[0107] 3) Obtain the prior location

[0108] Take out the element with index index_prior from the prior location list and pass its coordinates to an instance node_rand of the Node class.

[0109] 4) Find the nearest node

[0110] Use the nearest_neighbor method to find the node node_near closest to node_rand. self.vertex may be a list or set containing all nodes.

[0111] 5) Define the target point and the boundaries of the sampling area

[0112] target_local is the coordinate of the target point, and bounds_local defines the boundaries of the sampling area. Here, the boundaries are a rectangular area formed by the nearest node and the target point.

[0113] 6) Conduct goal-biased sampling

[0114] Use the goal_biased_sampling method to conduct goal-biased sampling within the defined sampling area. goal_bias_prob is the probability of goal-biased sampling, provided by self.direct_prob. The sampling result sample is used to create a new node node_new.

[0115] 7) Return the result

[0116] The method returns the prior location prior_pos, the nearest node node_near, and the newly sampled node node_new.

[0117] For the flowchart of the single key point-guided sampling method 1, see Figure 6 .

[0118] For prior sampling within a specific local area, that is, for target-biased sampling and generating new sampling points based on double key points and local scale parameters.

[0119] Function parameters of the double key point-guided sampling method 2:

[0120] self: Usually refers to an instance of a class, indicating that this is a class method.

[0121] prior_pos: Prior position, representing the double key points (latitude, longitude, and altitude) for reference during sampling.

[0122] node_near: The node closest to a certain target point.

[0123] local_scale: Local scale parameter, used to control the range of the sampling area.

[0124] Method implementation:

[0125] 1) Initialize the new node list

[0126] Create an empty list node_new_list to store the newly generated nodes.

[0127] 2) Loop for sampling

[0128] Loop local_scale times, and generate a new sampling point in each loop.

[0129] 3) Calculate the new local target position

[0130] Calculate a scaling factor scale based on the current loop count fruit and local_scale, and then use this factor, the prior position prior_pos, and the coordinates of the nearest node node_near to calculate the new local target position target_local_new.

[0131] 4) Find the node closest to the new target

[0132] Use the nearest_neighbor method to find the node closest to the new target position target_local_new and update node_near.

[0133] 5) Define the sampling range

[0134] Define the sampling range bounds_local, that is, the ranges of x and y, which are the coordinates of node_near and target_local_new respectively. As mentioned above, the height parameter z is omitted in the usage example of the task template.

[0135] 6) Perform goal-biased sampling

[0136] Use the goal_biased_sampling method to perform goal-biased sampling within the defined range to generate a new sampling point sample. goal_bias_prob is the probability parameter for goal-biased sampling.

[0137] 7) Add the new node to the list

[0138] Convert the generated sampling point sample into a Node object and add it to the node_new_list list.

[0139] 8) Return the list of new nodes

[0140] Return the list node_new_list containing all newly generated nodes.

[0141] Supplement to the implementation of the above path planning method:

[0142] 1) nearest_neighbor and goal_biased_sampling are methods defined in two related codes and need to be implemented in advance. nearest_neighbor is used to find the nearest node.

[0143] 2) goal_biased_sampling is used to perform goal-biased sampling within the specified range.

[0144] 3) self.vertex is a list or data structure containing all nodes.

[0145] 4) self.direct_prob is the probability parameter for goal-biased sampling, which is used to control the degree of bias towards the goal during sampling.

[0146] 5) This method assumes that Node is a custom class containing x and y attributes to represent the coordinates (latitude and longitude) of the node. As mentioned above, the height parameter z is omitted in the usage example of the task template.

[0147] For the flowchart of the double key-point guided sampling method 2 or strategy, see Figure 7

[0148] The evaluation experiment adopts a synthetic cost. By comprehensively considering distance, height, energy consumption, and obstacle distance, an RRT-star cost function that meets the actual requirements is designed, and the selection of weight coefficients needs to be adjusted according to specific application scenarios. The synthetic cost is:

[0149] ;

[0150] Among them, h1 and h2 are the height values of nodes q1 and q2 respectively; E(h1, h2) is the energy consumption function caused by height change, which is usually proportional to the height difference; in the obstacle distance penalty term, d(q1, q2) is the distance between the trajectory and the nearest obstacle; λ1, λ2, and λ3 are weight coefficients used to balance the influence of each term.

[0151] (3) The MEC Server feeds back an updated task template to the UAV-Agent.

[0152] Due to the complex environment of low-altitude cities, the environmental perception and target recognition information of the UAV-Agent may be biased, airspace restrictions may be updated in real time, and new navigation visual language inputs may occur. Therefore, calibration and update by the MEC Server are required.

[0153] (4) The UAV-Agent downloads the updated task template to the navigation station.

[0154] Through the UAV-Agent, the navigation station will obtain a low-altitude UAV navigation knowledge base.

[0155] (5) After the UAV-Agent receives the ACK sent by the navigation station, it leaves the service area involved in the geofence.

[0156] The UAV-Agent cruises in several geofences of the urban grid to download the low-altitude UAV navigation knowledge base to the distributed navigation stations.

[0157] Figure 1 The key lies in demonstrating the information transfer process among various entities and how to cooperate to complete the construction of the low-altitude UAV navigation knowledge base. This is necessary for understanding the working mechanism of the UAV navigation system and the lightweight interaction among its components.

[0158] Such as Figure 2 As shown, for the information transfer process among the user machine, navigation station, UAV-Agent, and MEC Server during the navigation task, the four vertical boundary lines respectively represent the user machine, navigation station, UAV-Agent, and MEC Server. There are multiple arrows between each boundary line, indicating different message transfer processes.

[0159] (1) The user machine at the starting point entering the geofence sends a navigation task request to the navigation station. Here, the user machine has a device with low-level visual language navigation function and can send navigation task requests in different forms, such as instructions, text, or even voice memos.

[0160] (2) After receiving the request, the navigation station selects the guiding path data from the task template and feeds back the navigation path to the user machine. When the navigation station collects real-time congestion data, it will perform graph calculations to obtain a sub-optimal navigation path. When the navigation task request of the user machine is a text or even a voice prompt, the navigation station uses the local embodied intelligence module to understand the navigation task request. Thus, the navigation station is a device with medium or high-level visual language navigation capabilities.

[0161] (3) After receiving the guiding path data, the user machine starts the subsequent process of autonomous navigation. It sequentially passes through the path key points on the guiding path and uses low-level control to fine-tune the (virtual) lane within 10 meters to cope with three-dimensional space disturbances, such as static ground buildings, wires and branches moving with the wind, and surrounding drones.

[0162] (4) When the current navigation task ends, the user machine notifies the navigation station and the navigation station feeds back ACK to the user machine.

[0163] (5) Finally, the UAV-Agent communicates with the MEC Server, stores the path trajectory, and updates the task template, etc.

[0164] Figure 2 It shows the information transfer process among the user machine, the navigation station, the UAV-Agent, and the MEC Server, and illustrates how the low-altitude UAV and its ground autonomous navigation vehicle share information with the navigation station through a wireless interface. The dynamic interaction between the navigation station and the UAV-Agent (when ID2 takes 4) is not reflected here, and it is optional. This is important for understanding the working process and communication mechanism of the autonomous navigation system.

[0165] When ID2 takes 1, 2, 3, and 4, the involved UAV control system utilizes the already constructed UAV navigation knowledge base. When ID2 takes 3, the key points are also called prior positions. With the assistance of other components or sensors (such as cameras, lidars, and millimeter-wave radars, etc.), the navigation station will be responsible for coordinating the navigation and path planning of the UAV within the geofence. The geofence may contain one or multiple navigation stations, and the UAV is not specifically located within the visual range of the navigation station.

[0166] In the path planning experiment, a class is used to define the environmental parameters and obstacles. By setting the coordinate ranges related to longitude and latitude, x_range = [-20, 250] (meters) and y_range = [-40, 40] (meters), and the height z_range is taken from 15 to 45 meters. Secondly, the boundary obstacles, rectangular obstacles, and circular obstacles are initialized through the obs_boundary, obs_circle, and obs_rectangle functions to configure the environmental constraint conditions required for path planning. The geofence takes rectangles of multiple different heights. See Figure 3 , in addition to the boundary obstacles simulating the geofence, each rectangular area in the benchmark test includes one or more rectangular obstacles. The navigation station is located between two rectangular obstacles and is not within the line of sight of the starting point.

[0167] The following is divided into two paragraphs to further explain and supplement the task template involved in Table 1 and Figure 4 and Figure 5 the graph structure involved. Table 1 is a semantic entity suitable for LLMs and Agents to understand the semantics of the navigation scenario and involves the data collection and processing process. The graph structure is a data structure convenient for modeling three-dimensional (virtual) roads and lanes. The latter uses heterogeneous nodes to model the three-dimensional airspace of the city, thus facilitating the implementation of multi-layer path planning at different heights. The UAV-Agent is responsible for constructing the first task template, and the semantic entity it outputs is the initial entity. However, the low-altitude flight management system considered can dynamically update the task template based on mobile edge computing. This task template and graph structure are downloaded to the navigation station. The navigation station can not only perform graph calculations but also perform RRT-star simulations. As mentioned above, the guiding path in Table 1 contains several path key point parameters, but the nodes of the graph structure are not directly equivalent to the path key points in Table 1. See Figure 3 In the two-dimensional plane in , the midpoint of the left boundary of a region is just one of the path key points. The guiding path involved in Table 1 is only a static and initial set of path planning parameters, which do not necessarily contain the coordinates of a large number of points in the trajectory and do not consider the kinematic constraints and dynamic obstacles during local range navigation. In short, the need for a safe and smooth path is inseparable from the RRT-star algorithm, and the guiding path is only a sampling of the trajectory output by RRT-star. Secondly, the definition of the initial graph structure requires the use of the parameters in Table 1. After performing RRT-star, the edge weights of the initial graph structure will be reset to the edge weights of the new graph structure.

[0168] Continue with the above description and supplement. Using the key points of the guiding path in Table 1, the navigation station adopts the RRT-star algorithm for path planning. The purpose of setting the guiding path in the task template is for planning across multiple regions. Therefore, the task template focuses on semantic regional guidance, the graph structure is convenient for evaluating multi-layer path planning to adapt to dynamic low-altitude traffic, while RRT-star focuses on local navigation planning. Since the boundaries of the two regions may be different, the traditional RRT-star algorithm is sensitive to boundary parameters and is prone to falling into local traps at the boundaries of the two regions. At this time, using the prior reference points provided by the task template, another goal-biased RRT-star algorithm can be realized, that is, a sampling strategy that crosses the boundaries of the two regions.

[0169] Summary of the above two paragraphs: For the low-altitude flight management system in the low-altitude economy, the present invention innovatively designs semantic entities, graph structure implementation and goal-biased RRT-star, and makes the three stages complement each other.

Claims

1. A method for guiding the three-dimensional low-altitude urban operation of an unmanned aerial vehicle based on a lightweight task template, characterized in that, It includes the following steps: (1) The UAV intelligent agent UAV-Agent enters the geofence and sends a starting point collection instruction to the navigation station within the geofence; (2) After receiving the starting point collection instruction, the navigation station feeds back the starting point, ending point information within the geofence and the correct reception information ACK to the UAV-Agent; (3) The UAV-Agent plans an initial path to run from the starting point to the ending point according to the feedback information, collects the environmental information within the geofence during the running process, and constructs an initial task template based on the collected environmental information; And updates the task template through the mobile edge server; The task template includes a graph structure obtained by rasterizing a three-dimensional low-altitude area, several path key points, and several guiding paths planned according to the graph structure and path key points; (4) Send the updated task template to the navigation station. The navigation station uses a path planning algorithm for the path key points on the guiding path in the task template to perform path planning on the guiding path and obtain a smooth navigation path; the path planning algorithm includes an improved RRT-star algorithm; the improved RRT-star algorithm assists the sampling of sampling points in the RRT-star algorithm through path key points. Specifically: Sequentially use the path key points on the guiding path as prior positions; For the interval between the starting point and the path key point, obtain the prior position, find the node closest to the prior position, confirm the coordinates of the target point, and use the rectangular area formed by the closest node and the target point as the sampling area, and perform target-biased sampling within the sampling area to sample a new node; For the interval between two path key points, calculate a new local target position through a scaling factor, prior position, and coordinates of the closest node, find the node closest to the new target position, and use the rectangular area formed by the node closest to the new target position and the local target position as the sampling area, and perform target-biased sampling within the sampling area to sample a new node; (5) The user machine at the starting point within the geofence sends a task request to the navigation station; (6) After receiving the task request, the navigation station selects the guiding path data from the task template and feeds back the navigation path to the user machine; (7) After receiving the guiding path data and the navigation path, the user machine starts the subsequent process of autonomous navigation and sequentially passes through the path key points on the guiding path; (8) When the current navigation task ends, the user machine notifies the navigation station and the navigation station feeds back ACK to the user machine.

2. The method for guiding the three-dimensional low-altitude urban operation of an unmanned aerial vehicle based on a lightweight task template according to claim 1, wherein The task template includes the task type and corresponding process, the longitude, latitude, and altitude of the navigation station, the geofence boundary, the longitude, latitude, and altitude of the starting point and ending point within the geofence, the number of guiding paths, the longitude, latitude, and altitude of several path key points, several guiding paths, the set of area IDs after rasterizing the three-dimensional low-altitude area, the area ID index, and the area topological relationship.

3. The method for guiding the three-dimensional low-altitude urban operation of an unmanned aerial vehicle based on a lightweight task template according to claim 2, wherein The construction steps of the graph structure in the task template are: (31) Divide the three-dimensional low-altitude urban structure within the geofence into multiple grids with geometric topological relationships; (32)Map the grid into multiple virtual roads, abstract a virtual road as a graph structure of isomorphic nodes, and store a single path, the path key points on the path, and graph calculations related to path planning in the graph structure of each isomorphic node; the graph structure describes the connection relationships of different regions in the environment; virtual roads at different heights are abstracted as graph structures of heterogeneous nodes, and the graph structure of heterogeneous nodes includes several paths, and the height of each path is different. (33)Assign weights to the edges of the graph structure.

4. The method for guiding the three-dimensional low-altitude urban operation of an unmanned aerial vehicle based on a lightweight task template according to claim 3, characterized in that Calculate the synthesis cost of the trajectory planned by the improved RRT-star algorithm, and update the weights of the edges in the graph structure through the synthesis cost; the synthesis cost The calculation formula is as follows: ； Among them, is the height value of node , is the height value of node , is the energy consumption function caused by height change, is the distance between the trajectory and the nearest obstacle, , and are weight coefficients.

5. The method for guiding the three-dimensional low-altitude urban operation of an unmanned aerial vehicle based on a lightweight task template according to claim 4, wherein The path planning methods adopted by the initial path in step (3) include: A* algorithm, Dijkstra algorithm, and RRT algorithm. The environmental information collected during the operation includes image acquisition, location information request, and geographic API call to obtain data of path key points. After the UAV-Agent reaches the end point, it generates an initial task template from the collected information and uploads it to the database of the mobile edge server (MEC server) at the same time. The MEC server calibrates and updates the initial task template, and feeds back the updated task template to the UAV-Agent. The UAV-Agent downloads the updated task template to the navigation station. After the UAV-Agent receives the ACK sent by the navigation station, it leaves the service area involved in the geographical fence.

6. The method for guiding the three-dimensional low-altitude urban operation of an unmanned aerial vehicle based on a lightweight task template according to claim 5, wherein The task template also includes spatial limit boundaries, traffic data, and meteorological data.

7. The method for guiding the three-dimensional low-altitude urban operation of an unmanned aerial vehicle based on a lightweight task template according to claim 6, wherein In step (6), after the navigation station receives the task request, when the navigation station collects real-time congestion degree data, the navigation station will perform graph calculations to obtain a sub-optimal guiding path.

8. A computer device, comprising a memory, a processor, and a computer program stored on 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 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 7.

Images