Multi-unmanned aerial vehicle cooperative exploration method for autonomously focusing high-value area

Abstract

The invention discloses a multi-unmanned aerial vehicle cooperative exploration method for autonomously focusing a high-value area. The method comprises the following steps: S1, environment modeling and map updating: establishing an environment module and dynamically updating a global map; s2, boundary detection and task generation: based on hierarchical spatial decomposition and boundary sensing clustering, constructing a boundary detection and exploration task generation mechanism, and screening out a high-value region containing rich geometric information; s3, distributed task allocation: based on a pairwise interaction mechanism and a vehicle path problem with capacity constraint, performing distributed allocation on the tasks generated in the step S2, so that the unmanned aerial vehicle autonomously decides without a central control station; s4, path planning and execution: performing multi-level planning and trajectory optimization based on coverage path guidance, and controlling the unmanned aerial vehicle to execute; according to the method and the device, the problems of low exploration resource utilization rate, low exploration path efficiency and poor environmental adaptability in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the field of autonomous exploration by multiple unmanned aerial vehicles (UAVs), and in particular to a method for collaborative exploration by multiple UAVs that autonomously focus on high-value areas. Background Technology

[0002] As a prime example of the deep integration of artificial intelligence and general aviation technology, unmanned aerial vehicle (UAV) systems have become an important platform for achieving intelligent environmental perception and autonomous decision-making. In missions such as disaster relief, industrial inspection, and military reconnaissance, UAVs need to rapidly and comprehensively explore unknown environments without prior maps. With increasing mission complexity, single UAVs face significant bottlenecks in exploration efficiency, coverage, and system robustness, making collaborative exploration by multiple UAVs an inevitable development trend.

[0003] Existing autonomous exploration methods for unmanned aerial vehicles (UAVs) are mainly divided into two categories: front-edge detection-based and sampling-based planning. The former uses the front boundary as the exploration target, which has high computational efficiency but lacks global path optimization and is prone to repeated exploration; the latter evaluates information gain through random sampling, which can obtain better paths but has high computational overhead and is difficult to meet the requirements of real-time tasks. In recent years, in order to balance efficiency and global optimization, some studies have introduced multi-vehicle path planning problem modeling, and achieved collaborative exploration through spatial decomposition and task allocation.

[0004] However, existing methods generally suffer from the following technical defects: First, the lack of a space exploration value assessment mechanism treats high-information-density areas such as obstacle boundaries as the same as open and unknown areas, resulting in unreasonable allocation of exploration resources; second, insufficient adaptability to environmental structures, failing to fully consider the impact of obstacle distribution on exploration paths, leading to a significant decrease in efficiency in structured environments; and third, low system resource utilization, with a large amount of computing, communication, and energy resources consumed in low-value areas, restricting the improvement of overall exploration performance.

[0005] In-depth analysis reveals that the root cause of the shortcomings in existing technologies lies in the lack of effective utilization of environmental semantic information. Specifically, the spatial distribution of obstacles is a key factor determining the value of exploration. In this invention, we define areas near obstacle surfaces that provide more information about environmental structure as "high-value areas," for which existing methods lack quantitative assessment and response mechanisms. This inherent deficiency prevents UAV systems from achieving targeted and optimized allocation of exploration resources, resulting in poor performance in large and complex environments.

[0006] Therefore, there is an urgent need for an autonomous multi-UAV exploration method capable of intelligently selecting high-value exploration areas based on environmental structural characteristics and optimizing resource allocation to overcome existing technological bottlenecks. An ideal solution should possess environmental structure perception capabilities, dynamically assess exploration value through obstacle spatial constraints, and concentrate limited onboard resources on exploring key areas, thereby achieving a qualitative improvement in exploration efficiency in complex scenarios. This invention aims to address the aforementioned technical deficiencies by providing a novel method that autonomously focuses on high-value exploration areas based on obstacle spatial constraints, significantly improving the exploration efficiency and resource utilization of multi-UAV systems. Summary of the Invention

[0007] This invention addresses the problems of low resource utilization, low exploration path efficiency, and poor environmental adaptability in existing technologies for multi-UAV exploration, and proposes a multi-UAV collaborative exploration method that autonomously focuses on high-value areas.

[0008] The technical solution of the present invention is as follows:

[0009] A multi-UAV collaborative exploration method that autonomously focuses on high-value areas includes the following steps:

[0010] Step S1, Environment Modeling and Map Update: Establish environment modules and dynamically update the global map;

[0011] Step S2, Boundary Detection and Task Generation: Based on hierarchical spatial decomposition and boundary-aware clustering, a boundary detection and exploration task generation mechanism is constructed;

[0012] Step S3, Distributed Task Allocation: Based on the pairwise interaction mechanism and the vehicle routing problem with capacity constraints, the tasks generated in step S2 are distributed and allocated in a distributed manner, enabling the UAV to make autonomous decisions without a central control station.

[0013] Step S4, Path Planning and Execution: Perform multi-level planning and trajectory optimization based on the covered path guidance, and control the UAV to execute.

[0014] Preferably, step S1 includes the following steps:

[0015] Step S11: Perceive the environment and collect data using airborne sensors;

[0016] Step S12: Fuse sensor data to build and update the full-grid map;

[0017] Step S13: Based on the raster map, construct the Euclidean Signed Distance Field (ESDF).

[0018] Preferably, in step S11, the functions of the environment module include local environment perception and global map management.

[0019] Preferably, step S2 includes the following steps:

[0020] Step S21: Online maintenance and updating of the hierarchical mesh used for spatial decomposition;

[0021] Step S22: Detect and validate valid boundary voxels representing the boundaries of unexplored regions in the hierarchical mesh;

[0022] Step S23: Use the region growing algorithm to cluster the spatially continuous effective boundary voxels into exploration regions of balanced size;

[0023] Step S24: Around each of the exploration areas, the system samples multiple candidate viewpoints;

[0024] Step S25: Perform multi-dimensional evaluation and ranking of the candidate viewpoints, including information gain and travel cost, to select the target viewpoint;

[0025] Step S26: Establish an obstacle distance filtering mechanism based on ESDF to ensure the safety of the target viewpoint.

[0026] Preferably, step S3 includes the following steps:

[0027] Step S31: Construct a pairwise interaction mechanism for communication and mission negotiation among multiple UAVs;

[0028] Step S32: Model the task allocation problem as a vehicle pathing problem with capacity constraints, where the UAV is the vehicle, its exploration capability is the capacity, and the target viewpoint is the city;

[0029] Step S33: Solve the vehicle path problem with capacity constraints and extract an independent task sequence path for each UAV;

[0030] Step S34: Generate a sparse task graph representing the global task allocation relationship based on the task sequence path.

[0031] Preferably, step S4 includes the following steps:

[0032] Step S41: Based on the sparse task graph, plan a global coverage path for each UAV that covers all its assigned tasks;

[0033] Step S42: Guided by the global coverage path, perform local viewpoint optimization and fine-grained obstacle avoidance path planning;

[0034] Step S43: Using the minimum time objective, generate a smooth and dynamically feasible B-spline trajectory;

[0035] Step S44: During trajectory execution, replanning is performed based on real-time environmental information to ensure collision avoidance;

[0036] Step S45: Track and control the drone to execute the final generated trajectory.

[0037] Compared with existing technologies, the beneficial effects of the multi-UAV collaborative exploration method for autonomously focusing on high-value areas in this invention are as follows:

[0038] 1. This invention establishes an intelligent exploration value assessment system based on obstacle spatial constraints, designs an exploration value quantification method based on Euclidean symbolic distance field, develops a correlation model between obstacle distance and information gain to achieve priority differentiation of different spatial regions, and constructs a multi-criteria assessment framework that comprehensively considers spatial proximity, information potential, and exploration cost.

[0039] 2. This invention achieves targeted optimization of exploration resources, dynamically filters high-value exploration targets through obstacle distance constraints to avoid resource waste, designs a load-aware task allocation mechanism to ensure a balanced distribution of workload among UAVs, and develops an adaptive allocation strategy for computing resources to prioritize the planning needs of key areas.

[0040] 3. This invention enhances the system's ability to perceive and adapt to environmental structures, enables online adjustment of intelligent exploration strategies based on environmental geometric features, and develops a spatial decomposition method for structure perception, so that task division matches environmental features.

[0041] 4. This invention breaks through the efficiency bottleneck of existing technologies, significantly improves the information acquisition efficiency of unit resources through a value-driven exploration strategy, establishes a scalable distributed architecture, supports efficient collaboration of large-scale UAV swarms, and achieves near-global optimal exploration path planning while maintaining communication efficiency.

[0042] 5. The core innovative objective of this invention is to fundamentally solve the problem of the insensitivity of existing technologies to the value density of spatial information by introducing an exploration value assessment mechanism guided by obstacle spatial constraints, thereby achieving breakthrough progress in the efficiency and practicality of multi-UAV (Unmanned Aerial Vehicle) collaborative exploration. Attached Figure Description

[0043] Figure 1(a) is a simulation experiment of a UVA according to the present invention.

[0044] Figure 1(b) is a simulation experiment diagram of a UVA according to the present invention.

[0045] Figure 2(a) is a simulation experiment diagram of two UVAs of the present invention.

[0046] Figure 2(b) shows the simulation experiment of the two UVAs of the present invention.

[0047] Figure 3(a) shows the simulation experiment of the four UVAs of the present invention.

[0048] Figure 3(b) shows the simulation experiment of four UVAs of the present invention.

[0049] Figure 4(a) shows the simulation experiment of the six UVAs of the present invention.

[0050] Figure 4(b) shows the simulation experiment of the six UVAs of the present invention.

[0051] Figure 5(a) shows the simulation experiment of the eight UVAs of the present invention.

[0052] Figure 5(b) shows the simulation experiment of eight UVAs of the present invention.

[0053] Figure 6(a) shows the simulation experiment of ten UVAs of the present invention.

[0054] Figure 6(b) shows the simulation experiment of ten UVAs of the present invention.

[0055] Figure 7 This is a flowchart of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and examples.

[0057] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0058] A multi-UAV collaborative exploration method that autonomously focuses on high-value areas includes the following steps:

[0059] Step S1, Environment Modeling and Map Update: Establish environment modules and dynamically update the global map;

[0060] Step S2, Boundary Detection and Task Generation: Based on hierarchical spatial decomposition and boundary-aware clustering, a boundary detection and exploration task generation mechanism is constructed;

[0061] Step S3, Distributed Task Allocation: Based on the pairwise interaction mechanism and the vehicle routing problem with capacity constraints, the tasks generated in step S2 are distributed and allocated in a distributed manner, enabling the UAV to make autonomous decisions without a central control station.

[0062] Step S4, Path Planning and Execution: Perform multi-level planning and trajectory optimization based on the covered path guidance, and control the UAV to execute.

[0063] Step S1 of this implementation plan includes the following steps:

[0064] Step S11: Perceive the environment and collect data using airborne sensors;

[0065] Step S12: Fuse sensor data to build and update the grid map;

[0066] Step S13: Based on the raster map, construct the Euclidean Signed Distance Field (ESDF).

[0067] In step S11 of this implementation scheme, the functions of the environment module include local environment perception and global map management.

[0068] Step S2 of this implementation plan includes the following steps:

[0069] Step S21: Online maintenance and updating of the hierarchical mesh used for spatial decomposition;

[0070] Step S22: Detect and verify valid boundary voxels representing the boundaries of unexplored regions in the hierarchical grid; Step S23: Use a region growing algorithm to cluster spatially continuous valid boundary voxels into exploration regions of balanced size;

[0071] Step S24: Around each of the exploration areas, the system samples multiple candidate viewpoints;

[0072] Step S25: Perform multi-dimensional evaluation and ranking of the candidate viewpoints, including information gain and travel cost, to select the target viewpoint;

[0073] Step S26: Establish an obstacle distance filtering mechanism based on ESDF to ensure the safety of the target viewpoint.

[0074] Step S3 of this implementation plan includes the following steps:

[0075] Step S31: Construct a pairwise interaction mechanism for communication and mission negotiation among multiple UAVs;

[0076] Step S32: Model the task allocation problem as a vehicle path problem with capacity constraints, where the UAV is the vehicle, its exploration capability is the capacity, and the target viewpoint is the city; Step S33: Solve the vehicle path problem with capacity constraints and extract an independent task sequence path for each UAV; Step S34: Based on the task sequence path, generate a sparse task graph representing the global task allocation relationship.

[0077] Preferably, step S4 includes the following steps:

[0078] Step S41: Based on the sparse task graph, plan a global coverage path for each UAV that covers all its assigned tasks; Step S42: Guided by the global coverage path, perform local viewpoint optimization and fine obstacle avoidance path planning; Step S43: Using a minimum time objective, generate a smooth and dynamically feasible B-spline trajectory; Step S44: During trajectory execution, replan based on real-time environmental information to ensure collision avoidance; Step S45: Track and control the UAV to execute the finally generated trajectory.

[0079] When this implementation plan is implemented,

[0080] This invention proposes a multi-UAV collaborative exploration method that autonomously focuses on high-value areas. The core idea of ​​this method is to introduce an obstacle distance constraint mechanism to intelligently select exploration areas with high information gain, thereby avoiding the waste of exploration resources in open and unknown spaces without obstacles, and thus achieving efficient and balanced collaborative exploration of multiple UAVs in complex environments.

[0081] Specifically, this method first constructs an environment model through distributed volume mapping and hierarchical spatial decomposition, and uses Euclidean symbolic distance field to calculate obstacle distances in real time. In the boundary detection stage, only valid boundary areas close to obstacles are retained, while "air boundaries" far from obstacles are excluded. Then, distributed task allocation is achieved through pairwise interaction and vehicle pathing problems with capacity constraints to ensure balanced load for each UAV. Finally, obstacle distance filtering is introduced in coverage path planning to generate coverage paths that only access high-value areas, and safe and agile exploration flights are achieved through local viewpoint optimization and minimum time trajectory generation.

[0082] The core steps of the technical solution.

[0083] Step 1: Environment Modeling and Map Update

[0084] (1) Overall Functional Description

[0085] The system described in this invention constructs an environment modeling and map updating mechanism based on distributed volume mapping and hierarchical spatial decomposition in multi-UAV collaborative autonomous exploration missions. This module aims to ensure real-time perception and consistent maintenance of the global map for multiple UAVs in communication-constrained environments, providing accurate environmental geometric information for subsequent decision-making and planning.

[0086] (2) Environmental perception and data acquisition

[0087] Each drone is equipped with onboard sensors such as depth cameras, LiDAR, or inertial measurement units to collect real-time depth information of the surrounding environment and its own pose. Through sensor fusion algorithms, the observation data in the drone's body coordinate system is transformed into the global coordinate system, generating a local sparse point cloud containing spatial geometric information, which serves as the underlying data source for map updates.

[0088] (3) Creation of raster maps

[0089] The system uses a 3D occupancy grid map as the basic representation of the environment. Each voxel... There are three states: known free area, obstacle area, and unknown area. To handle sensor noise and dynamically update the map state, the system uses a recursive Bayesian filter to update the occupancy probability of each voxel. Specifically, the occupancy probability is represented in log-odds form to ensure the stability of numerical calculations.

[0090]

[0091] in, Indicates from the initial time to Sensor observation sequence at time points, This is the prior logarithmic probability, usually set to 0. Real-time observation data is then processed using a raycasting model. It is integrated into the global map to achieve incremental updates of the environmental state.

[0092] (4) Construction of the Euclidean symbolic distance field

[0093] To support subsequent trajectory optimization and viewpoint selection based on obstacle distance constraints (ODC), the system constructs an ESDF in real time based on the occupied grid map. This invention employs the Fast Incremental Euclidean Distance Fields (FIESTA) algorithm to maintain the ESDF. For any voxel in the map... Its ESDF value This represents the Euclidean distance from the center of the voxel to the surface of the nearest obstacle. The update process follows the principle of wavefront propagation.

[0094]

[0095] in, It is a voxel The neighborhood set. This distance field not only provides accurate obstacle distance information. It also provides distance gradients for obstacle avoidance planning. .

[0096] (5) Multi-UAV map fusion and consistency maintenance

[0097] To address the global consistency issue among multiple UAVs under conditions of incomplete communication, this system employs a decentralized map fusion strategy based on map blocks and a bookkeeping mechanism.

[0098] Map tile data structure: The system divides the global map into several independent map data tiles. Each data tile contains metadata (such as index ID, timestamp, version number) and a differentially compressed voxel state sequence.

[0099] Communication Protocol: The system uses the User Datagram Protocol (UDP) to broadcast newly observed map data blocks, thereby avoiding the handshake overhead during connection establishment and significantly improving the real-time performance of data transmission.

[0100] Bookkeeping mechanism and synchronization: Each drone maintains a bookkeeping set in real time. Used to record in detail the map tile indexes it holds. Latest update timestamp and status version number The drone periodically broadcasts its bookkeeping set to neighboring nodes. By comparing its local records with the received bookkeeping information, the receiver quickly identifies missing or expired data and requests retransmission accordingly, thereby achieving consistent synchronization of the global map while minimizing communication bandwidth usage.

[0101] Conflict resolution: When multiple updates for the same area are received, the system adopts the "timestamp priority" principle, overwriting the old data with the latest observation data; at the same time, a spatial consistency check is performed to ensure that the merged map is not contradictory in physical geometry.

[0102] (6) Hierarchical grid decomposition mechanism

[0103] As a bridge connecting environment modeling and task allocation, the system maintains a hierarchical grid (Hgrid) structure. Based on distance information provided by ESDF, this structure dynamically decomposes the unknown space into cells of different levels and removes invalid regions far from obstacles, providing efficient spatial indexing support for the subsequent "Step 2: Boundary Detection and Task Generation".

[0104] Step 2: Boundary Detection and Task Generation

[0105] (1) Overall Functional Description

[0106] This step constructs a task generation mechanism based on hierarchical spatial decomposition and boundary-aware clustering. Its core objective is to intelligently identify truly valuable "valid frontiers" in the environment based on global map updates, and to eliminate "air boundaries" far from environmental structures through obstacle distance constraint (ODC), thereby generating a high-quality set of exploration tasks.

[0107] (2) Online maintenance and updating of hierarchical grids

[0108] Based on the global map established in step 1, the system maintains an HGrid structure in real time as a spatial index for boundary detection. When a map update area is received... At that time, the system triggers the dynamic update process of HGrid.

[0109] HGrid dynamic update mechanism:

[0110] 1. Initial state: When the environment is completely unknown. Only the coarsest level is included. The cell;

[0111] 2. Triggering condition: When a map update area is received. When this happens, the HGrid update process is initiated;

[0112] 3. Boundary constraints: For each Must meet ;

[0113] 4. Detailed Judgment: For each If with any Overlapping and satisfying If so, then proceed with uniform subdivision;

[0114] 5. Removal Conditions: When Or the cell centroid is unreachable or At that time, from Remove from;

[0115] (3) Effective boundary voxel detection and verification

[0116] The system incrementally detects valid boundary voxels within active HGrid cells.

[0117] Definition of an effective boundary voxel: a voxel A voxel is marked as a valid boundary voxel if and only if it simultaneously satisfies the following three conditions:

[0118] State conditions: It is in a known free state and there is at least one unknown voxel in its 26-neighborhood;

[0119] Sensor accessibility: It is within the effective detection range of the sensor.

[0120] Obstacle Distance Constraint (ODC): The distance to the nearest obstacle satisfies This constraint directly eliminates invalid boundary points that are suspended in the center of an open area.

[0121] (4) Boundary-aware region growing clustering

[0122] To transform discrete voxels into executable task objectives, the system employs a region growing algorithm to cluster spatially continuous effective boundary voxels into equally sized exploration regions (clusters). The specific process is as follows:

[0123] Input: Set of valid boundary voxels Current map M, ESDF

[0124] Output: Set of valid boundary clusters

[0125] 1. Initialize the access marker array

[0126] 2. For each unvisited boundary voxel :

[0127] a. If ,jump over

[0128] b. Initialize the new cluster ,queue

[0129] c.

[0130] d. When Non-empty and :

[0131] i. Retrieve the current voxel

[0132] ii.

[0133] iii. For each neighbor :

[0134] if and And the space is connected

[0135]

[0136] e. If and

[0137] 3. Return

[0138] (5) Multi-criteria viewpoint sampling and generation

[0139] Candidate viewpoints are systematically sampled around each valid boundary cluster, ensuring that viewpoints are close to the boundary and have practical exploration value. Uniform sampling strategy: for boundary clusters. At its center of mass Centered at a radius of 1, Uniform sampling within the range One perspective:

[0140]

[0141] in, ; Pointing to the cluster centroid.

[0142] Sampling constraints:

[0143] 1. Spatial feasibility: The viewpoint must be located in known free space;

[0144] 2. Boundary proximity: ;

[0145] 3. Sensor Accessibility: Verify the accessibility of the viewpoint through ray projection;

[0146] 4. Observation integrity: Ensure that the viewpoint can effectively observe the target boundary cluster;

[0147] 5. Safety: Distance between viewpoint and obstacle .

[0148] (6) Viewpoint quality assessment and multi-criteria ranking

[0149] In order to select the optimal target from the candidate viewpoints, the system establishes a multi-dimensional evaluation system that includes frontier coverage, unknown area coverage, and travel cost, and performs dual priority ranking.

[0150] Information gain metric:

[0151] Effective boundary coverage metric:

[0152]

[0153] in, It is a set of boundary voxels that has been verified for boundary validity.

[0154] Spatial expansion gain:

[0155]

[0156] Boundary distance score:

[0157]

[0158] Overall Exploration Value:

[0159]

[0160] Multi-level sorting mechanism: The system uses the following lexicographical order criterion to sort candidate viewpoints: Primary priority: based on the number of effective boundary coverages ( Arranged in descending order. The system prioritizes viewpoints that allow for observation of more environmental structural outlines to ensure the geometric integrity of the map construction.

[0161] Second priority: when When they are the same, they are ranked by the number of unknown areas covered. Sort in descending order. In this case, prioritize viewpoints with a wider field of view as supplementary options.

[0162] Prioritization: If both of the above are the same, then prioritize based on overall value. Sort.

[0163] (7) Obstacle distance screening mechanism

[0164] To ensure the feasibility of the generated exploration objectives in actual execution, a secondary screening based on obstacle distance is required for the sorted viewpoints. Dual threshold screening criteria: viewpoint It shall be reserved if and only if:

[0165]

[0166] in:

[0167] Minimum safe distance (e.g., 0.5m) to ensure the drone does not collide with obstacles;

[0168] The maximum effective distance threshold (e.g., 3.5m) is used to completely eliminate invalid "airborne viewpoints" (i.e., low-value targets) far from obstacles, thereby ensuring that the drone only locks onto and focuses on high-value areas near obstacles for exploration, forcing the drone to perform its mission within a safe corridor close to the environmental surface.

[0169] Step 3: Distributed Task Allocation and Load Balancing

[0170] This step aims to construct a distributed task allocation system based on a pairwise interaction mechanism and the Capacitated Vehicle Routing Problem (CVRP). Its core objective is to endow each UAV with independent decision-making capabilities in a distributed architecture without a central node, achieving dynamic and balanced allocation of exploration tasks among multiple UAVs through local communication alone, effectively avoiding the communication bottlenecks and single-point-of-failure risks inherent in traditional centralized allocation methods.

[0171] 1. Paired interaction mechanism

[0172] Each drone The system periodically attempts to negotiate tasks with neighboring drones within its communication range. To ensure efficient and fair interaction, the system employs the following protocol flow:

[0173] Neighbor Discovery: Obtains a set of currently communicable neighboring drones via broadcast communication. .

[0174] Conflict Avoidance: Inspect Neighbor's Drone Is it recently? Interactions attempted within the specified time (by comparing current time) and of If so, skip that neighbor.

[0175] Fairness guarantee: Select the neighbor with the longest time interval since the last successful interaction, i.e.:

[0176]

[0177] Request-Response Protocol

[0178] Request phase: Towards Send a request message containing the reassigned task set. and and the current timestamp . set up And wait for a response.

[0179] Response phase: After receiving the request, check if you have recently... Interaction attempted within a short period of time If so, the request is rejected. Otherwise, Accept the request and update your task set to and set and . Successful response sent to .

[0180] Confirmation phase: After receiving a successful response, update your task set to... and set If no response is received within the timeout period, the interaction will be abandoned.

[0181] 2. CVRP Problem Modeling

[0182] 2.1 Problem Definition

[0183] Given two drones and and their current task set (HGrid cell collection). and Merge task sets into ,Include One cell.

[0184] Objective: To Reclassified into and , so that:

[0185]

[0186]

[0187] Minimize the total length of the coverage path of the two drones.

[0188] Satisfying the capacity constraint: The volume of unknown space allocated to each drone does not exceed the total volume of unknown space. Proportion.

[0189] 2.2 Graph Model Construction

[0190] Build a containing A directed graph with nodes:

[0191] Each node corresponds to an HGrid cell.

[0192] The two nodes correspond to the current location of the drone.

[0193] 1 virtual warehouse node

[0194] Node index allocation:

[0195] Virtual warehouse: Node 0

[0196] drones Node 1

[0197] drones Node 2

[0198] HGrid cell: Node 3 to node

[0199] 2.3 Definition of Cost Matrix

[0200] Cost Matrix The structure is as follows:

[0201]

[0202] in:

[0203] It is a very large positive number (e.g.) );

[0204] Cost from drones to HGrid cells;

[0205] Cost between HGrid cells.

[0206] Cost between HGrid cells:

[0207]

[0208] in, It is a cell The center of mass, From arrive Collision-free path It is the path length.

[0209] 3. CVRP Solving and Path Extraction

[0210] 3.1 Problem Solving

[0211] The above CVRP problem is solved using the LKH-3 solver (Lin-Kernighan-Helsgaun). LKH-3 transforms the CVRP into an equivalent Traveling Salesman Problem (TSP) and handles capacity constraints using a penalty function.

[0212] 3.2 Path Extraction

[0213] The solver returns two paths (one for each drone). Each path starts in the virtual warehouse, connects to a drone, visits a series of HGrid cells, and finally returns to the virtual warehouse.

[0214] When extracting the actual coverage path, ignoring the virtual repository node, we get:

[0215] drones Coverage path: from Begin by accessing the HGrid cells assigned to it in sequence.

[0216] drones Coverage path: from Begin by accessing the HGrid cells assigned to it in sequence.

[0217] 4. Sparse graph

[0218] To provide efficient data structure support for subsequent fine-grained path planning, the system constructs a sparse task graph based on the assigned task sequence. In this graph, nodes represent only the assigned high-value HGrid cells, and edges connect only spatially adjacent cells with collision-free paths. Compared to fully connected graphs, the sparse task graph significantly reduces the search space and computational overhead of subsequent global path planning, enabling the system to maintain high efficiency in large-scale environments.

[0219] Step 4: Path and Trajectory Planning

[0220] (1) Overall Functional Description

[0221] The system described in this invention constructs a hierarchical planning mechanism based on coverage path guidance in multi-UAV collaborative autonomous exploration missions, decomposing the global task into three levels: coverage path, local path, and trajectory optimization. Its goal is to generate smooth, efficient, and collision-free flight trajectories while ensuring safety and dynamic feasibility, guiding UAVs to quickly complete exploration tasks. The entire process requires no ground station intervention, achieving fully autonomous planning and execution. This module consists of a coverage path planning unit, a local viewpoint optimization unit, and a trajectory generation unit, realizing the transformation from task allocation to specific flight maneuvers. The core improvement of this invention lies in introducing a region filtering mechanism based on obstacle spatial constraints in the coverage path planning stage, ensuring that UAVs only explore areas near boundaries with high information gain, avoiding wasting resources in unobstructed, unknown spaces.

[0222] (2) Coverage path planning

[0223] Coverage Path (CP) provides each drone with a global order of access to its assigned HGrid cells, ensuring that the drone scans high-value portions of its area of ​​responsibility in an efficient manner. Unlike traditional methods, this invention dynamically filters cells worth exploring by evaluating obstacle proximity in real time.

[0224] Overlay path generation steps:

[0225] 1. Input: The collection of HGrid cells currently assigned to the drone. The current location of the drone .

[0226] 2. Cell filtering based on obstacle constraints: Calculate the distance from the centroid of each cell to the nearest obstacle, and filter according to a preset exploration distance threshold to obtain a set of high-value cells.

[0227] 3. Construct a collision-free path graph between cells: For the filtered set of cells, use a sparse graph to calculate the shortest path length between the centroids of the cells.

[0228] 4. Model as a Traveling Salesman Problem (TSP): Starting from the drone's current position, visit the centroids of all cells and return a visit sequence that minimizes the total path length.

[0229] 5. Solve the TSP: Use the LKH-3 solver to obtain the coverage path. ,in It is a cell The center of mass.

[0230] Dynamic update mechanism:

[0231] When task assignment results are updated or map changes cause cell states to change, the coverage path is recalculated. A rolling planning approach is used, executing only the first part of the coverage path. Each cell was reorganized.

[0232] (3) Local viewpoint optimization Under the guidance of the global path, the system adopts the retceding horizontal planning strategy to perform fine viewpoint optimization only for local tasks within the current field of view.

[0233] Local path planning steps:

[0234] 1. Input: the path to be covered Each cell, and the set of boundary clusters contained within those cells. .

[0235] 2. Problem Modeling: Construct a model based on the current viewpoint of the drone. Starting from the first point, visit in sequence The boundary cluster in the middle, and finally reach the first Centroid of each cell The path planning problem.

[0236] 3. Graph construction: The candidate viewpoints of each boundary cluster are used as graph nodes, and the edge weights are the transition times between viewpoints (considering changes in movement and yaw).

[0237] 4. Solve for TSP with fixed start and end points: Solve using a method similar to step (2) to obtain the local path. .

[0238] 5. For each boundary cluster, select one from multiple candidate viewpoints to minimize the transition time and maximize the coverage gain of the entire path.

[0239] 6. Use Dijkstra's algorithm on the graph to search for the optimal viewpoint sequence.

[0240] 7. Viewpoint selection optimization:

[0241] For each boundary cluster, select one from multiple candidate viewpoints to minimize the transition time and maximize the coverage gain of the entire path.

[0242] 8. Use Dijkstra's algorithm on the graph to search for the optimal viewpoint sequence.

[0243] (4) Minimum Time B-Spline Trajectory Generation To transform the discrete viewpoint sequence into a continuous and smooth control trajectory, the system uses a uniform B-spline to parameterize the path. The trajectory is represented by a uniform B-spline curve:

[0244]

[0245] in, It is a control point. yes B-spline basis functions.

[0246] Trajectory optimization problem: optimizing control points Minimize the following objective function:

[0247]

[0248] The costs are as follows:

[0249] Smoothness cost (acceleration, jerk, etc.)

[0250] Total trajectory time

[0251] Collision cost with obstacles

[0252] Collision costs with other drones

[0253] Cost of velocity and acceleration constraints

[0254] Boundary state cost (matching the start and end states)

[0255] (5) Real-time replanning and collision avoidance: To ensure safety, the system supports real-time replanning during trajectory execution.

[0256] Replanning trigger conditions: discovery of new obstacles (ESDF update makes the area traversed by the trajectory unsafe); conflict with the predicted trajectory of other drones or excessive tracking error.

[0257] Replanning strategy: While keeping the global objective unchanged, the system regenerates obstacle avoidance trajectories only within a local area. If no feasible solution can be found, an emergency hover is executed and global planning is retried.

[0258] (6) Trajectory tracking and control

[0259] After trajectory generation, trajectory tracking is performed using the underlying geometry controller or model predictive controller (MPC). Position controller: based on the trajectory... Calculate the desired thrust vector. Attitude controller: Calculates the desired attitude based on the yaw angle setting and acceleration direction.

[0260] Through the above hierarchical planning mechanism, the system can achieve efficient and safe autonomous exploration in complex and unknown environments, while ensuring collaborative operations among multiple UAVs.

[0261] Table 1 shows a comparison of data from 10 simulated flights of a UVA in different scenarios.

[0262] Table 1 Comparison of the three methods in two scenarios

[0263]

[0264] Table 2 shows a comparison of data from multiple drones, including data from 10 flights in different scenarios.

[0265] 1, 2, 4, 6, 8, 10 represent the number of drones, and Racer represents the existing technology solution.

[0266] Table 2 Comparison of data from multiple drones in different scenarios

[0267]

Claims

1. A multi-UAV collaborative exploration method for autonomously focusing on high-value areas, characterized in that, Includes the following steps: Step S1, Environment Modeling and Map Update: Establish environment modules and dynamically update the global map; Step S2, Boundary Detection and Task Generation: Based on hierarchical spatial decomposition and boundary-aware clustering, a boundary detection and exploration task generation mechanism is constructed; Step S3, Distributed Task Allocation: Based on the pairwise interaction mechanism and the vehicle routing problem with capacity constraints, the tasks generated in step S2 are distributed and allocated in a distributed manner, enabling the UAV to make autonomous decisions without a central control station. Step S4, Path Planning and Execution: Perform multi-level planning and trajectory optimization based on the covered path guidance, and control the UAV to execute.

2. The multi-UAV collaborative exploration method for autonomously focusing on high-value areas according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Perceive the environment and collect data using airborne sensors; Step S12: Fuse sensor data to build and update the grid map; Step S13: Construct the Euclidean symbolic distance field based on the grid map.

3. The multi-UAV collaborative exploration method for autonomously focusing on high-value areas according to claim 2, characterized in that, In step S11, the functions of the environment module include local environment perception and global map management.

4. The multi-UAV collaborative exploration method for autonomously focusing on high-value areas according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Online maintenance and updating of the hierarchical mesh used for spatial decomposition; Step S22: Detect and verify valid boundary voxels representing the boundaries of unexplored regions in the hierarchical grid; Step S23: Use a region growing algorithm to cluster spatially continuous valid boundary voxels into exploration regions of balanced size; Step S24: Around each of the exploration areas, the system samples multiple candidate viewpoints; Step S25: Perform multi-dimensional evaluation and ranking of the candidate viewpoints, including information gain and travel cost, to select the target viewpoint; Step S26: Establish an obstacle distance filtering mechanism based on the Euclidean symbolic distance field to ensure the safety of the target viewpoint.

5. The multi-UAV collaborative exploration method for autonomously focusing on high-value areas according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Construct a pairwise interaction mechanism for communication and mission negotiation among multiple UAVs; Step S32: Model the task allocation problem as a vehicle path problem with capacity constraints, where the UAV is the vehicle, its exploration capability is the capacity, and the target viewpoint is the city; Step S33: Solve the vehicle path problem with capacity constraints and extract an independent task sequence path for each UAV; Step S34: Generate a sparse task graph representing the global task allocation relationship based on the task sequence path.

6. The multi-UAV collaborative exploration method for autonomously focusing on high-value areas according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Based on the sparse task graph, plan a global coverage path for each UAV that covers all its assigned tasks; Step S42: Guided by the global coverage path, perform local viewpoint optimization and fine obstacle avoidance path planning; Step S43: Using the minimum time objective, generate a smooth and dynamically feasible B-spline trajectory; Step S44: During trajectory execution, replan based on real-time environmental information to ensure collision avoidance; Step S45: Track and control the UAV to execute the finally generated trajectory.

Images