Agent-based drone route planning method

Abstract

This invention relates to an agent-based unmanned aerial vehicle (UAV) flight path planning method, comprising: discretizing the UAV flight area using a three-dimensional mesh to obtain a set of cubes, and obtaining the semantic representation vector of each cube in the set; parsing and decomposing the user's flight path planning task instructions through an agent to obtain a sequence of semantic subtasks, and obtaining the semantic representation vector of each subtask in the sequence; matching the subtasks with the semantic representation vectors of the cubes to obtain a target cube set for each subtask; finally, using a designed semantic A* search algorithm to search for flight paths, generating flight path segments that satisfy each subtask; and assembling the generated flight path segments into a complete three-dimensional flight path according to the order of the subtasks. This method can automatically generate refined flight paths that meet semantic intent and geometric safety requirements, realizing end-to-end intelligent planning from semantic instructions to executable flight paths.

Description

Technical Field

[0001] This invention relates to the field of path planning technology, and more particularly to a method for planning flight routes for unmanned aerial vehicles (UAVs). Background Technology

[0002] Traditional drone route planning typically relies solely on geometric obstacle avoidance (digital elevation models, buildings, obstacles, etc.) and cannot automatically plan routes based on task semantics (such as "inspecting three water accumulation points near a building" or "inspecting along the Qingyuan River").

[0003] The typical method used in existing UAV flight path planning is as follows: First, relying on digital elevation models, oblique photogrammetry models, or laser point cloud data, a pure geometric 3D model of the terrain, buildings, and other obstacles in the task area is constructed. Then, the planning system receives a series of spatial coordinate points or regions manually preset by the operator (e.g., building vertices selected on the 3D model, polylines drawn on a map), and treats them as waypoints or coverage areas that must be reached. Finally, the planning algorithm, using these manually preset spatial points as constraints, combined with the UAV's physical performance parameters and a fixed safe obstacle avoidance distance, calculates the shortest or optimal energy-efficient flight trajectory in 3D geometric space that connects all waypoints and avoids the obstacle model. The core of this method lies in completely transforming the complex task into a sequence of manually predefined spatial points, and then having the algorithm perform pure geometric path calculations.

[0004] However, the aforementioned existing methods have significant drawbacks. The fundamental problem lies in the complete disconnect between the planning process and the task semantics. Because the system can only process geometric information such as coordinates, distance, and shape, it cannot understand higher-level semantic concepts such as "factory roof," "high-voltage line," and "greenway." This leads to a high reliance on operators to perform tedious and precise manual spatial translation and annotation for specialized tasks, resulting in low automation. Therefore, existing technologies struggle to respond to users' higher-level semantic commands, and the planning results only meet geometric obstacle avoidance requirements, exhibiting significant deficiencies at the task functionality level and limiting flexibility and practicality. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method for planning unmanned aerial vehicle (UAV) routes that can be planned according to the semantic instructions of the user, in contrast to the above-mentioned prior art.

[0006] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a method for planning flight routes of unmanned aerial vehicles (UAVs), characterized by the following steps: Step 1: Discretize the UAV flight area to obtain a set of cubes represented by a 3D mesh; and obtain the interest point data of the UAV flight area. Based on the interest point data, obtain the semantic representation vector of each cube in the cube set. Step 2: The intelligent agent parses and decomposes the user's route planning task instructions to obtain a semantic subtask sequence, which includes semantic description texts of multiple subtasks; the semantic description texts of each subtask are input into the semantic embedding model to obtain the semantic representation vector of each subtask. Step 3: Match the semantic representation vector of the subtask in Step 2 with the semantic representation vector of the cube in Step 1 to obtain the target cube set for each subtask. In addition, initialize the semantic workload of each target cube in the target cube set for the current subtask to obtain the initial semantic workload of each target cube for the current subtask. Step 4: Execute each subtask in the order from front to back, generate route segments that satisfy each subtask, and assemble the generated route segments into a complete three-dimensional route according to the order of the subtasks. The specific process of executing the j-th subtask is as follows: the starting point of the current route segment, the target cube set of the j-th subtask, and the initial semantic workload of each target cube in the target cube set of the j-th subtask for the j-th subtask are used as inputs to the designed semantic A* search algorithm. The semantic A* search algorithm is used to search for routes and generate route segments that satisfy the j-th subtask. The semantic A* search algorithm is designed based on the A* search algorithm and a semantic heuristic function is added to the heuristic function of the A* search algorithm.

[0007] To mark each cube for subsequent flight path planning, step 1 also includes the following processing: acquiring digital elevation model data and 3D scene data of the UAV flight area, and judging each cube in the cube set based on the digital elevation model data and 3D scene data to identify obstacle cubes and risk cubes, and assigning a mark to each cube in the cube set. If any cube is an obstacle cube or a risk cube, then a no-fly sign is assigned to that cube; otherwise, a flyable sign is assigned to that cube.

[0008] Preferably, the method for determining obstacle cubes and risk cubes is as follows: if the spatial extent of any cube overlaps with the digital elevation model data or 3D scene data, then that cube is determined to be an obstacle cube; each obstacle cube is expanded, and the cubes generated by the expansion of obstacle cubes are determined as follows: if the spatial extent of a cube generated by the expansion of obstacle cubes overlaps with the digital elevation model data or 3D scene data, then that cube is determined to be a risk cube. By expanding each obstacle cube, a safety buffer zone is introduced within the neighborhood of the obstacle cube.

[0009] Preferably, the interest point data in step 1 includes the spatial location of the interest points and the corresponding semantic text. The specific process for obtaining the semantic representation vector of each cube in the cube set is as follows: Step 1-1: Obtain the set of interest points associated with each cube in the cube set; Steps 1-2: Input the semantic text of each interest point into the semantic embedding model to generate the semantic representation vector of each interest point; Steps 1-3: Calculate the semantic representation vector of each cube; Cubes in a set of cubes semantic representation vector The calculation formula is: ; in, For points of interest For cube semantic contribution weights ; For points of interest The semantic representation vector; For a cube The associated set of points of interest, It is an empty set.

[0010] Preferably, the specific process of obtaining the set of interest points associated with each cube in step 1-1 is as follows: For any interest point, determine the influence domain of the interest point on the horizontal plane. The influence domain of the interest point on the horizontal plane is a circular area centered on the spatial position of the current interest point on the horizontal plane, and the radius of the circular area is a preset value; For any cube, if a cube falls within the influence domain of an interest point on the horizontal plane in the horizontal projection plane, it is determined that the cube has a spatial association with the current interest point; Obtain the set of interest points associated with each cube in the above manner.

[0011] Preferably, the specific process of obtaining the target cube set for each sub-task in step 3 is as follows: Step 3-1: Calculate the similarity between the semantics of the subtask and the semantics of each cube according to the following formula; ; in, For the first Semantics of Subtasks and Cubes The similarity between semantics; For the first Semantic description text for each subtask For the first The semantic representation vectors of each subtask The vector cosine similarity function; Step 3-2: Filter the cubes as follows to obtain the target cube set for each subtask; The selection method is as follows: if the similarity between the semantics of the current subtask and the semantics of a certain cube is greater than the preset similarity threshold, then the current cube is taken as the target cube corresponding to the current subtask; if the target cube set of the current subtask is empty, then the maximum similarity between the semantics of the current subtask and the semantics of each cube is obtained, and the cube corresponding to the maximum similarity is taken as the target cube corresponding to the current subtask.

[0012] Preferably, the initial semantic workload of each target cube in the target cube set in step 3 for the current subtask is initialized as follows: calculate the similarity between the semantics of the current subtask and the semantics of a target cube in the target cube set of the current subtask, and use the similarity as the initial semantic workload of the target cube for the current subtask.

[0013] Preferably, in step 4, when executing the first sub-task, the cube where the UAV takes off is taken as the starting point of the current flight segment; when executing the second to J sub-tasks, where J is the total number of sub-tasks, the cube where the end point of the flight segment of the previous sub-task is taken as the starting point of the current flight segment.

[0014] Specifically, the heuristic function in the semantic A* search algorithm The calculation formula is: ; in, For distance heuristic function; For semantic heuristic functions; The calculation formula is: ; For semantic terms, weight coefficients The gating coefficient, , Neighboring cube The nearest distance to the target set of cubes, neighboring cubes The next step for drones is to fly in all directions around a cube. For gating scale parameters; The cube where the drone is currently located Transfer to neighboring cube The total amount of remaining semantic work at that time This represents the initial remaining semantic workload for the current subtask. , The set of target cubes for the current subtask. for The target cube in For the target cube The initial semantic workload for the current subtask.

[0015] Specifically, The calculation formula is: ; in, To start from the cube where the drone is currently located Transfer to neighboring cube At that time, the remaining semantic workload of each target cube, , For drones in the neighborhood cube The visible intersection target subset in The cube where the drone is currently located Target cube The remaining semantic workload, .

[0016] Compared with existing technologies, the advantages of this invention are as follows: By using an intelligent agent to parse and decompose the user's flight path planning task instructions, no manual conversion of spatial coordinates is required, significantly reducing the operational threshold and professional dependence, and improving planning efficiency. Furthermore, the initial semantic workload of each target cube for the current sub-task is used as input to the designed semantic A* search algorithm, which searches for flight paths to generate flight path segments that satisfy the sub-task. The semantic A* search algorithm is based on the A* search algorithm, with a semantic heuristic function added to the heuristic function of the A* search algorithm. This method can automatically generate refined flight paths that meet semantic intent and geometric safety requirements, achieving end-to-end intelligent planning from semantic instructions to executable flight paths. This significantly enhances the accuracy, adaptability, and practicality of autonomous UAV operations in complex scenarios, providing an efficient and intelligent solution for various UAV operations. Attached Figure Description

[0017] Figure 1 This is a flowchart of the UAV route planning method in an embodiment of the present invention. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0019] like Figure 1 As shown, the UAV route planning method in this embodiment includes the following steps: Step 1: Discretize the UAV flight area to obtain a set of cubes represented by a 3D mesh; and obtain the interest point data of the UAV flight area. Based on the interest point data, obtain the semantic representation vector of each cube in the cube set. Step 2: The intelligent agent parses and decomposes the user's route planning task instructions to obtain a semantic subtask sequence, which includes semantic description texts of multiple subtasks; the semantic description texts of each subtask are input into the semantic embedding model to obtain the semantic representation vector of each subtask. Step 3: Match the semantic representation vector of the subtask in Step 2 with the semantic representation vector of the cube in Step 1 to obtain the target cube set for each subtask. In addition, initialize the semantic workload of each target cube in the target cube set for the current subtask to obtain the initial semantic workload of each target cube for the current subtask. Step 4: Execute each subtask in the order from front to back, generate route segments that satisfy each subtask, and assemble the generated route segments into a complete three-dimensional route according to the order of the subtasks.

[0020] The interest point data in step 1 includes the spatial location of the interest points and their corresponding semantic text. In this embodiment, the specific process of obtaining the semantic representation vector of each cube in the cube set is as follows: Step 1-1: Obtain the set of interest points associated with each cube in the cube set; Steps 1-2: Input the semantic text of each interest point into the semantic embedding model to generate the semantic representation vector of each interest point; Steps 1-3: Calculate the semantic representation vector of each cube; Cubes in a set of cubes semantic representation vector The calculation formula is: ; in, For points of interest For cube semantic contribution weights ; For points of interest The semantic representation vector; For a cube The associated set of points of interest, It is an empty set.

[0021] In this embodiment, the specific process of obtaining the set of interest points associated with each cube is as follows: For any interest point, determine the influence domain of the interest point on the horizontal plane. The influence domain of the interest point on the horizontal plane is a circular area centered on the spatial position of the current interest point on the horizontal plane, and the radius of the circular area is a preset value; For any cube, if a cube falls within the influence domain of an interest point on the horizontal plane in the horizontal projection plane, it is determined that the cube has a spatial association with the current interest point; Obtain the set of interest points associated with each cube in the above manner.

[0022] In this embodiment, the specific process of obtaining the target cube set for each sub-task in step 3 is as follows: Step 3-1: Calculate the similarity between the semantics of the subtask and the semantics of each cube according to the following formula; ; in, For the first Semantics of Subtasks and Cubes The similarity between semantics; For the first Semantic description text for each subtask For the first The semantic representation vectors of each subtask The vector cosine similarity function; Step 3-2: Filter the cubes as follows to obtain the target cube set for each subtask.

[0023] In this embodiment, the cube selection method is as follows: if the similarity between the semantics of the current subtask and the semantics of a certain cube is greater than a preset similarity threshold, then the current cube is taken as the target cube corresponding to the current subtask; if the target cube set of the current subtask is empty, then the maximum similarity between the semantics of the current subtask and the semantics of each cube is obtained, and the cube corresponding to the maximum similarity is taken as the target cube corresponding to the current subtask; when there are multiple cubes with the same or similar maximum similarity, the multiple cubes are included in the target cube set.

[0024] In this embodiment, the initial semantic workload of each target cube in the target cube set for the current subtask is initialized as follows: the similarity between the semantics of the current subtask and the semantics of a target cube in the target cube set of the current subtask is calculated, and the similarity is used as the initial semantic workload of the target cube for the current subtask.

[0025] The specific process of executing the j-th subtask in step 4 is as follows: The starting point of the current route segment, the target cube set of the j-th subtask, and the initial semantic workload of each target cube in the target cube set for the j-th subtask are used as inputs to the designed semantic A* search algorithm. The semantic A* search algorithm is then used to search for routes and generate route segments that satisfy the j-th subtask. The semantic A* search algorithm is based on the A* search algorithm, with a semantic heuristic function added to the heuristic function of the A* search algorithm. The aforementioned A* search algorithm is existing technology and will not be elaborated further here.

[0026] When executing the first subtask, the cube containing the UAV's takeoff point is taken as the starting point of the current flight path segment. When executing the second to Jth subtasks, where J is the total number of subtasks, the cube containing the end point of the previous subtask's flight path segment is taken as the starting point of the current flight path segment. If there are no empty paths in the complete 3D flight path, the complete 3D flight path is taken as the final planned flight path; if there are empty paths in the complete 3D flight path, the system outputs "Unable to plan this semantic flight path" and issues an alarm.

[0027] Heuristic function in semantic A* search algorithm The calculation formula is: ; in, For distance heuristic function; For semantic heuristic functions; the above Used to guide the cube where the current drone is located toward the target set of cubes; This is used to guide the search to consume the remaining semantic requirements as quickly as possible after the cube where the current drone is located approaches the set of target cubes; The calculation formula is: ; For semantic terms, weight coefficients The magnitude of the value affects the strength of the semantic heuristic's influence on the overall heuristic. Larger targets tend to "complete semantic work" and will actively move towards observable targets (perhaps taking a slightly roundabout route). If it is small, it leans more towards "geometric shortest", with only slight semantic guidance; The gating coefficient, , Neighboring cube The nearest distance to the target set of cubes, neighboring cubes The next step for drones is to fly in all directions around a cube. For gating scale parameters; The cube where the drone is currently located Transfer to neighboring cube The total amount of remaining semantic work at that time This represents the initial remaining semantic workload for the current subtask. , The set of target cubes for the current subtask. for The target cube in For the target cube The initial semantic workload for the current subtask; The calculation formula is: , To start from the cube where the drone is currently located Transfer to neighboring cube At that time, the remaining semantic workload of each target cube, , For drones in the neighborhood cube The visible intersection target subset in The cube where the drone is currently located Target cube The remaining semantic workload, .

[0028] Step 1 also includes the following processing steps: acquiring digital elevation model data and 3D scene data of the UAV flight area, and judging each cube in the cube set based on the digital elevation model data and 3D scene data to identify obstacle cubes and risk cubes, and assigning a flag to each cube in the cube set. If any cube is an obstacle cube or a risk cube, a no-fly flag is assigned to that cube; otherwise, a flyable flag is assigned to that cube. The cube's flag will be used in the semantic A* search algorithm.

[0029] In this embodiment, the three-dimensional scene data includes three-dimensional entity data such as buildings and structures; the method for determining obstacle cubes and risk cubes is as follows: if the spatial range of any cube overlaps with the digital elevation model data or the three-dimensional scene data, then the cube is determined to be an obstacle cube; each obstacle cube is expanded, and the cube generated by the expansion of the obstacle cube is judged as follows: if the spatial range of a cube generated by the expansion of the obstacle cube overlaps with the digital elevation model data or the three-dimensional scene data, then the cube is determined to be a risk cube.

[0030] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An agent-based unmanned aerial vehicle (UAV) route planning method, characterized in that... Includes the following steps: Step 1: Discretize the UAV flight area to obtain a set of cubes represented by a 3D mesh; and obtain the interest point data of the UAV flight area. Based on the interest point data, obtain the semantic representation vector of each cube in the cube set. Step 2: The intelligent agent parses and decomposes the user's route planning task instructions to obtain a semantic subtask sequence, which includes semantic description texts of multiple subtasks; the semantic description texts of each subtask are input into the semantic embedding model to obtain the semantic representation vector of each subtask. Step 3: Match the semantic representation vector of the subtask in Step 2 with the semantic representation vector of the cube in Step 1 to obtain the target cube set for each subtask. In addition, initialize the semantic workload of each target cube in the target cube set for the current subtask to obtain the initial semantic workload of each target cube for the current subtask. Step 4: Execute each subtask in the order from front to back, generate route segments that satisfy each subtask, and assemble the generated route segments into a complete three-dimensional route according to the order of the subtasks. The specific process of executing the j-th subtask is as follows: The semantic A* search algorithm is designed by taking the starting point of the current route segment, the target cube set of the j-th subtask, and the initial semantic workload of each target cube in the target cube set of the j-th subtask for the j-th subtask as inputs. The semantic A* search algorithm is used to search for routes and generate route segments that satisfy the j-th subtask. The semantic A* search algorithm is designed based on the A* search algorithm and a semantic heuristic function is added to the heuristic function of the A* search algorithm. Heuristic function in semantic A* search algorithm The calculation formula is: in, For distance heuristic function; For semantic heuristic functions; The calculation formula is: in, For semantic terms, weight coefficients The gating coefficient, , Neighboring cube The nearest distance to the target cube set, neighboring cubes The next step for drones is to fly in all directions around a cube. For gating scale parameters; The cube where the drone is currently located Transfer to neighboring cube The total amount of remaining semantic work at that time This represents the initial remaining semantic workload for the current subtask. , The set of target cubes for the current subtask. for The target cube in For the target cube The initial semantic workload for the current subtask.

2. The UAV route planning method according to claim 1, characterized in that: Step 1 also includes the following processing procedures: The system acquires digital elevation model data and 3D scene data of the UAV flight area, and judges each cube in the cube set based on the digital elevation model data and 3D scene data to identify obstacle cubes and risk cubes. The system then assigns a flag to each cube in the cube set. If any cube is an obstacle cube or a risk cube, a no-fly sign is assigned to that cube. Conversely, the cube is given a flight symbol.

3. The UAV route planning method according to claim 2, characterized in that: The method for determining obstacle cubes and risk cubes is as follows: If the spatial extent of any cube overlaps with digital elevation model data or 3D scene data, then the cube is identified as an obstacle cube. Each obstacle cube is expanded, and the cubes generated by the expansion of the obstacle cubes are judged as follows: if the spatial range of a cube generated by the expansion of the obstacle cube overlaps with the digital elevation model data or the 3D scene data, then the cube is judged as a risk cube.

4. The UAV route planning method according to claim 1, characterized in that: The interest point data in step 1 includes the spatial location of the interest points and their corresponding semantic text. The specific process for obtaining the semantic representation vector of each cube in the cube set is as follows: Step 1-1: Obtain the set of interest points associated with each cube in the cube set; Steps 1-2: Input the semantic text of each interest point into the semantic embedding model to generate the semantic representation vector of each interest point; Steps 1-3: Calculate the semantic representation vector of each cube; Cubes in a set of cubes semantic representation vector The calculation formula is: in, For points of interest For cube semantic contribution weights ; For points of interest The semantic representation vector; For a cube The associated set of points of interest, It is an empty set.

5. The UAV flight path planning method according to claim 4, characterized in that: The specific process for obtaining the set of interest points associated with each cube in step 1-1 is as follows: For any point of interest, determine the influence domain of the point of interest on the horizontal plane. The influence domain of the point of interest on the horizontal plane is a circular area with the current spatial position of the point of interest on the horizontal plane as the center. The radius of the circular area is a preset value. For any cube, if a cube falls within the influence domain of a point of interest on the horizontal projection plane, then the cube is determined to have a spatial relationship with the current point of interest. The set of interest points associated with each cube is obtained using the method described above.

6. The UAV route planning method according to claim 4, characterized in that: The specific process for obtaining the target cube set for each subtask in step 3 is as follows: Step 3-1: Calculate the similarity between the semantics of the subtask and the semantics of each cube according to the following formula; in, For the first Semantics of Subtasks and Cubes The similarity between semantics; For the first Semantic description text for each subtask For the first The semantic representation vectors of each subtask The vector cosine similarity function; Step 3-2: Filter the cubes as follows to obtain the target cube set for each subtask; The filtering method is as follows: if the similarity between the semantics of the current subtask and the semantics of a certain cube is greater than the preset similarity threshold, then the current cube will be taken as the target cube corresponding to the current subtask. If the target cube set of the current subtask is empty, then the maximum similarity between the semantics of the current subtask and the semantics of each cube is obtained, and the cube corresponding to the maximum similarity is taken as the target cube of the current subtask.

7. The UAV route planning method according to claim 6, characterized in that: The initialization method for the semantic workload of each target cube in the target cube set in step 3 for the current subtask is as follows: Calculate the similarity between the semantics of the current subtask and the semantics of a target cube in the target cube set of the current subtask, and use this similarity as the initial semantic workload of the target cube for the current subtask.

8. The UAV flight path planning method according to claim 6, characterized in that: In step 4, when executing the first subtask, the cube where the UAV takes off is taken as the starting point of the current flight segment; when executing the second to J subtasks, where J is the total number of subtasks, the cube where the end point of the flight segment of the previous subtask is taken as the starting point of the current flight segment.

9. The UAV flight path planning method according to any one of claims 1 to 8, characterized in that: The calculation formula is: in, To start from the cube where the drone is currently located Transfer to neighboring cube At that time, the remaining semantic workload of each target cube, , For drones in the neighborhood cube The visual intersection target subset in The cube where the drone is currently located Target cube The remaining semantic workload, .

Images