An unmanned aerial vehicle layered autonomous exploration method and system based on coverage path guidance
By constructing portal hierarchy graphs and connectivity graphs, and combining global coverage path planning and local viewpoint evaluation, the problem of low efficiency in UAV autonomous exploration was solved, and efficient autonomous exploration in complex environments was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SUPER ROBOT RESEARCH INSTITUTE (HUANGPU)
- Filing Date
- 2026-02-09
- Publication Date
- 2026-06-12
Smart Images
Figure CN122195022A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of autonomous exploration technology for unmanned aerial vehicles (UAVs), specifically to a hierarchical autonomous exploration method and system for UAVs based on coverage path guidance. Background Technology
[0002] Autonomous exploration refers to the task of mapping unknown environments using mobile robots. It is a fundamental component of various robotic applications, such as structure inspection, 3D reconstruction, underground navigation, and search and rescue operations. Due to the limited battery life of unmanned aerial vehicle (UAV) platforms, it is crucial to research efficient exploration planning techniques that can cover accessible space as quickly as possible.
[0003] Several methods have been proposed to improve exploration efficiency, such as boundary-based methods and Next-Best-View (NBV) methods. These typically employ a greedy strategy, selecting the next target based on immediate rewards. This short-sighted strategy ignores global information, leading to unnecessary revisits to already explored areas. To address this, existing solutions introduce global guidance methods that consider all potential views. However, these solutions still fall short of the exploration task's objective: mapping the entire unknown environment. Most existing global guidance focuses on visiting all boundary regions or subspaces, neglecting unknown areas beyond the boundaries. Since boundaries change dynamically during exploration, this global guidance often produces highly inconsistent intentions, resulting in back-and-forth movement and failing to reflect the exploration task's objective. Therefore, this discrepancy leads to inefficient exploration paths, frequently overlapping with already explored areas, thus reducing overall exploration efficiency.
[0004] In recent years, attempts have been made to bridge this gap by introducing Coverage Paths (CPs) covering the entire unexplored space into exploration planning. However, due to several limitations, the potential of CPs to improve exploration efficiency has not been fully realized. First, the CPs generated by these methods often rely on simple spatial decomposition, failing to adequately capture the connectivity and topology of the environment, potentially leading to unreasonable CPs. Second, many CP planning modules suffer from high online computational overhead due to the complexity of problem solving, hindering subsequent planners from responding promptly and replanning upon receiving the latest environmental information. Furthermore, even if CPs provide reasonable guidance, some unexplored areas may remain in certain local details (such as behind corners or in small compartments), forcing UAVs to detour back to previous areas to finish, forming a long tail of exploration. This long tail directly reduces exploration efficiency and completion rate. Finally, simple viewpoint evaluation models cannot effectively balance the relationship between exploring new information and efficient advancement, resulting in low exploration efficiency. Summary of the Invention
[0005] To overcome the defects and shortcomings of existing technologies, this invention provides a hierarchical autonomous exploration method and system for unmanned aerial vehicles (UAVs) based on coverage path guidance. Based on viewpoint evaluation models, connectivity graphs, and cleanup mechanisms, this invention solves the efficiency and robustness problems faced by existing autonomous exploration technologies in large and complex environments, and achieves fast, smooth, and efficient autonomous exploration flight in complex environments.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] This invention provides a hierarchical autonomous exploration method for unmanned aerial vehicles (UAVs) based on coverage path guidance, comprising the following steps:
[0008] Acquire UAV pose information and environmental observation data to construct environmental perception and 3D voxel maps;
[0009] Construct a portal hierarchy graph, where each node represents a connected subspace and each edge represents a passable connection between subspaces established through the portal, generating free and unknown regions;
[0010] By using the centers of free and unknown regions as nodes in the portal layer graph, a portal layer connectivity graph is constructed.
[0011] Global coverage path planning based on portal layer connectivity graph;
[0012] Based on the global coverage path order, the current target subspace is determined, a local exploration viewpoint target sequence is generated, and local exploration viewpoint planning is performed.
[0013] Conduct on-site corner cleaning and dispatch;
[0014] The target is searched based on the current local exploration viewpoint target sequence, a navigation path is generated, and the 3D voxel map is updated.
[0015] As a preferred technical solution, acquiring UAV pose information and environmental observation data to construct environmental perception and 3D voxel maps specifically includes:
[0016] Environmental observation data is acquired using airborne visual sensors on UAVs, and the perception range is modeled as a three-dimensional view frustum.
[0017] The exploration space is represented in voxel form using an occupied grid map, which is divided into multiple three-dimensional voxel units. The voxel states are divided into free voxels, occupied voxels, and unknown voxels.
[0018] The boundary region between free voxels and unknown voxels is taken as the frontier region.
[0019] As a preferred technical solution, a portal hierarchy diagram is constructed, specifically including:
[0020] The space to be explored is divided into multiple unit spaces. When it is determined that the 3D voxel map has been updated, the unit spaces are decomposed.
[0021] Within the decomposed cell space, perform three-dimensional connected component labeling on voxels;
[0022] Based on the labeling results of the three-dimensional connected components, voxels with the same label are aggregated to form regions, regions composed of safe and free voxels are defined as free regions, and regions composed of unknown voxels are defined as unknown regions.
[0023] For each region, the corresponding voxel geometric center is calculated as the initial region center, and the initial region center is corrected.
[0024] As a preferred technical solution, a portal layer connectivity graph is constructed, specifically including:
[0025] Use the center of the free region as the vertex and the connectivity between free regions as the edge;
[0026] Construct connected edges for the unknown region;
[0027] For a spatial cell that contains both free and unknown regions, perform a restricted A* search within the cell between the corresponding free and unknown nodes. If the cell is determined to be reachable, establish a portal edge.
[0028] As a preferred technical solution, global coverage path planning is performed based on the portal layer connectivity graph, specifically including:
[0029] Determine the set of target nodes for the coverage path, and use the incremental A* algorithm to search for the path on the portal layer connectivity graph to obtain the coverage path;
[0030] Calculate the cost of the coverage path and retain the K candidate coverage paths with the minimum cost;
[0031] For each candidate coverage path, perform an A* search on the voxel map to calculate the actual path cost;
[0032] Compare The path with the lowest cost is selected from the calculated results of each path as the final global coverage path guide route.
[0033] As a preferred technical solution, the cost of the coverage path is calculated, specifically expressed as follows:
[0034] ;
[0035] in, In the picture The shortest path calculated above, Indicates the current location of the drone. Indicates the center of the candidate viewpoint. This represents the edge-to-edge cost given by the path length A*.
[0036] As a preferred technical solution, an A* search is performed on the voxel map to calculate the true path cost, specifically expressed as follows:
[0037] ;
[0038] in, This represents the A* path distance calculated in free space. This represents the calculated A* path distance in the unknown space. As a penalty factor, To estimate the feasible segmented path, Indicates the current location of the drone. This represents K candidate coverage paths.
[0039] As a preferred technical solution, based on the global coverage path order, the current target subspace is determined, and local exploration viewpoint planning is performed, specifically including:
[0040] Within the target subspace, the leading edge region is clustered, and within each leading edge cluster, the space is sampled according to the sensor field of view to generate several candidate viewpoints;
[0041] Calculate viewpoint evaluation metrics and adaptive viewpoint scores;
[0042] The adaptive viewpoint scores are sorted to obtain a local exploration viewpoint target sequence, and local exploration viewpoint planning is performed based on the local exploration viewpoint target sequence.
[0043] As a preferred technical solution, a viewpoint evaluation index is calculated, and an adaptive viewpoint score is calculated, expressed as follows:
[0044] ;
[0045] ;
[0046] ;
[0047] ;
[0048] ;
[0049] ;
[0050] ;
[0051] in, , , , For each weight parameter, Represents viewpoint entropy gain. Indicates the number of voxels. For finite coverage rays, K is the total number of rays in the sampling direction. Indicates the visibility of the viewpoint. Represents the normalized distance. Indicates the current location of the drone. The shortest path distance to candidate viewpoint v Indicates from arrive The set of all feasible paths, Representing an edge The weighted length, It is a minor term to prevent division by zero. and These are the shortest and longest path distances in the candidate viewpoint set, respectively. This represents the small cluster penalty term. Indicates a frontier cluster, Indicates the number of unknown voxels. This represents the cluster size threshold.
[0052] The present invention also provides a layered autonomous exploration system for UAVs based on coverage path guidance, for implementing the above-mentioned layered autonomous exploration method for UAVs based on coverage path guidance, including: a data acquisition module, a modeling module, a portal hierarchy graph construction module, a portal layer connectivity graph construction module, a global coverage path planning module, a local exploration viewpoint planning module, and an exploration module;
[0053] The data acquisition module is used to acquire UAV pose information and environmental observation data;
[0054] The modeling module is used to construct environmental perception and three-dimensional voxel maps;
[0055] The portal hierarchy graph construction module is used to construct a portal hierarchy graph, where each node represents a connected subspace, each edge represents a passable connection between subspaces established through the portal, and free regions and unknown regions are generated.
[0056] The portal layer connectivity graph construction module is used to construct the portal layer connectivity graph by using the free region center and the unknown region center as nodes of the portal layer graph.
[0057] The global coverage path planning module is used to perform global coverage path planning based on the portal layer connectivity graph.
[0058] The local exploration viewpoint planning module is used to determine the current target subspace based on the global coverage path order, generate a local exploration viewpoint target sequence, and perform local exploration viewpoint planning.
[0059] The exploration module is used to search for targets based on the current local exploration viewpoint target sequence, generate navigation paths, and update the three-dimensional voxel map.
[0060] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0061] Based on viewpoint evaluation models, connectivity graphs, and cleanup mechanisms, this invention solves the efficiency and robustness problems faced by existing autonomous exploration technologies in large and complex environments, and realizes rapid, smooth, and efficient autonomous exploration flight in complex environments. Attached Figure Description
[0062] Figure 1 This is a schematic diagram of the overall implementation architecture of the UAV hierarchical autonomous exploration method based on coverage path guidance of the present invention;
[0063] Figure 2 This is a schematic diagram of the spatial voxel representation of the present invention;
[0064] Figure 3 This is a schematic diagram showing the results of spatial decomposition and connectivity graph construction during the exploration process of this invention;
[0065] Figure 4 This is a schematic diagram of the portal layer connectivity graph construction of the present invention;
[0066] Figure 5 This is a schematic diagram of the global coverage path planning of the present invention;
[0067] Figure 6 This is a schematic diagram of viewpoint sampling and evaluation in this invention;
[0068] Figure 7 This is a schematic diagram of the cleaning process of the corner cleaning scheduler of the present invention. Detailed Implementation
[0069] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0070] Example 1
[0071] In a limited three-dimensional space In this process, the autonomous exploration mission of the UAV involves establishing a complete and accurate 3D voxel map of the unknown environment using onboard visual sensors. The space is represented as a set of cubic voxels, and the occupancy probability P(v) of each voxel V is continuously updated, incrementally transforming the initial unknown space. Mapped to two parts: (Free space) and (Space Occupancy) The sensor used for localization and exploration is a forward-looking stereo camera with a limited field of view (FoV). It is modeled as a frustum defined by the horizontal FoV, vertical FoV, and sensing depth. Because the sensor's perception stops at a surface, some hollow or corner spaces cannot be mapped. These spaces are... This indicates that the autonomous exploration planner needs to drive the drone to perceive and cover unknown spaces in the shortest possible time. When the entire region has been searched, that is, when and At that time, the task is considered fully completed;
[0072] To achieve efficient autonomous exploration planning, this embodiment adopts a global-local hierarchical planning architecture. When new sensor observation data arrives, planning preprocessing is first performed, including spatial decomposition, connectivity graph construction, front extraction, viewpoint update, and small corner sweep scheduler judgment. Then, based on this information, hierarchical planning guided by coverage path is performed to generate global coverage route and local viewpoint sequence. Finally, the motion trajectory is output. The entire process is iterated and repeated continuously during the exploration process until there are no uncovered fronts in the environment.
[0073] This embodiment introduces a coverage path as a global guidance strategy. A coverage path is a route that passes through all unexplored areas, aiming to guide the UAV to visit various areas in a high-level order. To construct the coverage path, the environment is divided into several sub-regions based on reachability and connectivity, and representative locations of these sub-regions are extracted as target points for global planning. These target points include two categories: free areas (explored space areas containing the leading viewpoint) and unknown areas (completely unexplored space). Then, a connecting path is planned between these target points to minimize the total flight distance. Specifically, global route planning is modeled as a Sequential Ordering Problem (SOP), i.e., finding the shortest path starting from the current position, passing through all target points, and not requiring a return to the starting point. Compared to the traditional greedy method of selecting the next viewpoint one by one, the planner in this embodiment considers all areas to be visited at the global level, thereby avoiding backtracking or suboptimal cruising caused by local decisions.
[0074] After the global coverage path determines the general exploration order, the local planner selects the specific next viewpoint based on the global intent. The process is as follows: determine the target partition that the coverage path should currently visit; then, among all candidate front viewpoints within that partition, select the best viewpoint as the next target based on an adaptive viewpoint evaluation model. Since there may be multiple front viewpoints in a partition, local planning needs to optimize the access order of these viewpoints, which can be viewed as a local SOP sequencing problem. Due to the limited partition size, the local viewpoint sequence can be optimized through simple sorting (from high to low viewpoint scores). When switching to the next global partition, the above process is repeated. In this way, the global coverage path provides the macro-order, and local planning ensures that the front viewpoint access order aligns with the global intent. Figure 1 Hierarchical planning ensures efficient navigation routes. Compared to methods that rely solely on local greedy approaches, it more consistently follows global guidance, avoiding ineffective detours or deviations from the intended direction.
[0075] Once the next target viewpoint and several subsequent viewpoint sequences are selected, the planner generates a smooth, dynamically feasible B-spline trajectory that connects the current UAV position with the target viewpoint sequence, ensuring that the constraints of smoothness, safety, and quadrotor dynamics are met. When environmental information is updated during trajectory execution (e.g., new obstacles are discovered or unknown areas expand), the system will promptly interrupt and replan to ensure safety and efficiency.
[0076] like Figure 1 As shown, the specific steps of the UAV hierarchical autonomous exploration method based on coverage path guidance in this embodiment include:
[0077] S1: Constructing environmental perception and a 3D voxel map, specifically including:
[0078] S11: Acquire drone status and sensor information;
[0079] During flight, the UAV acquires its own position and attitude information in real time, including its current position coordinates and attitude information; at the same time, it acquires environmental observation data through airborne vision sensors. The vision sensors have a limited field of view and a maximum sensing distance, and their sensing range is modeled as a three-dimensional field of view (FoV).
[0080] S12: Initialization and updating of 3D voxel map;
[0081] like Figure 2 As shown, based on the acquired environmental observation data, an occupancy grid map is used to represent the exploration space in voxels, dividing the environment into regular three-dimensional voxel units. An occupancy probability is maintained for each voxel and continuously updated with new observation data, thus dividing the voxel state into free voxels. Occupying voxels and unknown voxels ;
[0082] S13: Frontier region extraction: such as Figure 2 As shown, in the voxel map, the boundary region between free voxels and unknown voxels is taken as the frontier region. The frontier region represents potential exploration targets and provides a basis for subsequent viewpoint generation and exploration decisions.
[0083] S2: Portal Layer Spatial Decomposition and Subspace Modeling;
[0084] A portal hierarchy diagram structure is constructed at the global planning level to enhance the ability to explore and make decisions about large spaces. The portal hierarchy diagram is a hierarchical abstraction of space based on the actual connectivity of the environment: nodes represent connected subspaces (such as rooms, corridor sections, and other relatively open areas), and edges represent the passable connections between these subspaces established through narrow passages (portals). Portals generally refer to local bottleneck areas such as doorways, door frames, and narrow corridor entrances, which divide the environment into several relatively independent subspaces. Using a portal hierarchy diagram as a global representation can naturally reflect the topological structure of the environment and help plan a reasonable global route: drones need to enter each subspace one by one to explore and travel between areas through portals, so the global path should roughly follow the portal connection relationship.
[0085] S21: Initial spatial unit division;
[0086] In the initial exploration phase, the entire exploration space is divided into several uniform unit spaces based on the size of the environment to be explored. The size of each spatial unit is proportional to the sensor's field of view, and it serves as the basic task unit for subsequent spatial decomposition.
[0087] S22: Incremental spatial decomposition trigger;
[0088] During the exploration process, whenever the voxel map is updated, the bounding box is updated. Intersecting cells will be further broken down into more refined regions. Based on the latest raster occupancy information, separate non-connected regions, such as... Figure 3 As shown, the results of spatial decomposition and connectivity graph construction during the exploration process are obtained. The surrounding transparent cells do not intersect with the bounding box of the updated map and remain unchanged. The two areas highlighted in yellow are connected by a narrow, impassable corridor and are therefore divided into different areas. The rectangle in the middle right is an inaccessible hole, which is excluded from the exploration plan using the connectivity graph.
[0089] S23: Voxel connectivity-based region labeling;
[0090] Within each spatial cell that is triggered for decomposition, perform 3D connected component labeling on its voxels:
[0091] Besides the already occupied voxels and unknown voxels In addition to these two voxel states, according to free voxels Whether the minimum safe clearance distance is violated is further divided into safe free voxels and unsafe free voxels (because the size of the drone itself needs to be considered).
[0092] During the labeling process, only the unknown... and security and freedom Voxels are labeled, while occupied and unsafe free voxels are ignored.
[0093] Adjacent voxels are considered connected and assigned the same label only if both are safe free voxels or both are unknown voxels.
[0094] S24: Generation of free and unknown regions;
[0095] Based on the connectivity component labeling results, voxels with the same label are aggregated to form regions; regions composed of safe free voxels are defined as free regions, and regions composed of unknown voxels are defined as unknown regions. By ignoring unsafe free voxels, regions connected by impassable narrow passages are divided into different regions, such as... Figure 3 The two areas are highlighted in yellow.
[0096] S25: Calculation and correction of the regional center point;
[0097] For each region, its voxel geometric center is calculated as the initial region center, as follows:
[0098] ;
[0099] in, Represents a cell The center of a certain free region, This represents the voxel coordinates within the free region, where n represents the total number of voxels. Similarly, the cell can be calculated. Center of the unknown region .
[0100] However, since connectivity does not guarantee the convexity of a region, the initial center may fall into obstacles or other regions. In this case, the center needs to be corrected to the position of the nearest voxel within its region. As shown in Figure 3, the initial center is sampled within a circle of radius r. If the sampled point is located within this region, it is taken as the center of this region; if no sampled point meets the requirements, then... Expand the circle range to increase the step size, continue sampling, and check if there are any sampling points located within the region space; push inwards in turn until the center point of the region is found.
[0101] S3: Portal layer connectivity graph construction;
[0102] S31: Portal layer diagram node definition;
[0103] The free region centers and unknown region centers obtained in S2 are used as nodes in the portal layer graph;
[0104] In this context, nodes in the free region that contain the frontier viewpoint are marked as active free nodes, and nodes in the completely unknown region are marked as unknown nodes.
[0105] S32: Construction of connected edges in a free region;
[0106] After the portal layer graph is constructed, a connectivity graph is gradually built based on the constructed portal layer graph. In this context, the region center is used as a vertex, and the connectivity between regions is used as edges. It's important to note that this region-level connectivity differs from the inter-voxel connectivity mentioned earlier. For example... Figure 4 As shown, the connectivity graph consists of two subgraphs: and Capture free regions separately ( (area) and unknown area ( (region), and connections and edge ,Right now ;
[0107] For bounding box Each related cell The centers of the free region and the unknown region respectively constitute the vertices. and Edge construction is based on three types of restricted A* search, such as... Figure 4 As shown, for free subgraphs The edge in ,exist Each free vertex in and adjacent cells In The connectivity between the two cells is evaluated in pairs. This evaluation involves a restricted A* search in free space, which is confined to the region surrounding the two cells (e.g., ...). Figure 4 (a)). If a path is found, then the connection will be made. and Add the edge In the middle, the weight of the edge Defined as path length.
[0108] S33: Construction of connected edges in unknown regions;
[0109] Similarly, unknown subgraph The edge in Updated by performing a restricted A* search among vertices in the unknown voxel space (e.g.) Figure 4 (b)). If there is a feasible path between the unknown nodes, then establish a connecting edge between the unknown nodes.
[0110] S34: Portal Edge Construction;
[0111] For a spatial cell containing both free and unknown regions, a restricted A* search is performed within the cell between the corresponding free and unknown nodes; if reachable, a portal edge is established to describe the access relationship from the explored region to the unknown region (e.g., ...). Figure 4 (c)
[0112] S35: Unreachable subgraphs are removed;
[0113] After incremental construction, if the connectivity graph The presence of multiple unconnected components indicates the existence of unreachable voids in the environment, such as... Figure 3 As shown by the rectangle in the middle right, in this case, any isolated subgraph consisting only of unknown regions will be removed and excluded from the exploration plan.
[0114] This embodiment utilizes a portal layer graph to seek a global route that passes through all subspaces to be inspected. It finds the shortest path that traverses all specified nodes on the graph, which is equivalent to solving the path coverage problem of the graph. However, by abstracting the portal layer graph, the problem size is greatly reduced, and graph algorithms can be used to accelerate the computation.
[0115] S4: Global coverage path planning based on portal layer graph;
[0116] The global coverage path planning employs a two-stage mechanism of coarse estimation and fine calculation to balance efficiency and accuracy. First, an incremental A* algorithm is used to quickly search for coarsely estimated paths on the portal hierarchy graph. Then, several optimal candidates are finely evaluated. The coarse estimation stage ensures that when environmental information is only slightly updated, existing paths can be quickly updated without recalculation (because the coarsely estimated path calculation is segmented). The fine calculation stage, in the event of significant topological changes (changes in the global connectivity graph), performs a global A* path calculation on the k coarsely estimated candidate paths to prevent misleading coarse estimations from leading to suboptimal routes. Specifically, this includes:
[0117] S41: Determining the set of target nodes for the coverage path: Using all active free nodes and unknown nodes as the set of target nodes to be visited, construct a coverage path planning problem;
[0118] S42: Coverage path rough estimation stage: On the portal layer connectivity graph, the incremental A* algorithm is used to search for paths, and the previous planning result is used as the initial solution; a penalty factor is introduced for path segments that cross unknown areas to obtain an estimated coverage path that is biased towards the explored space;
[0119] Incremental A* coarse estimation is an algorithm suitable for finding shortest paths in dynamic graphs, efficiently updating paths when edge weights change or new nodes are added. In exploration scenarios, as the map expands, the weights of nodes and edges in the portal layer graph (e.g., the distance between two subspaces) are gradually updated. Using the previously planned global route as a starting point, when environmental updates cause graph changes, incremental A* is used to quickly adjust the estimated cost of the global route. The coarse cost estimation primarily considers the shortest path length between subspaces, which can be approximated by the Euclidean distance between portal centers. If some connections traverse unexplored areas, the unknown areas are temporarily treated as free space but a penalty factor is applied, resulting in a hypothetical path that might traverse the unknown and its cost, ultimately yielding the global route and distance estimates.
[0120] S43: Coverage Path Cost Estimation
[0121] For any portal edge, its rough cost is composed of the free space path length and the unknown space path length, where the unknown space path length is multiplied by a penalty factor; the current position of the UAV is mapped to the nearest free region node, which is used as the starting point of the coverage path;
[0122] Edges of any portal ,Door Door to door The rough estimate of the distance is:
[0123] ;
[0124] in, and Let represent the shortest paths in the free region and the unknown region, respectively. The length of the path segment traversing the unknown space is multiplied by a constant penalty factor. To account for the uncertainty of the unknown space, the portal layer edge weights are: ;
[0125] Anchor point mapping: Let the current position of the drone be... The candidate viewpoint center set is Connect the drone to its nearest viewpoint center, which in turn connects to a portal node that is the nearest unknown area center. :
[0126] ;
[0127] The edge-to-edge cost is given as A * path length:
[0128] ;
[0129] A rough estimate of a cross-portal path (from the current point to the viewpoint center) consists of three segments, as follows: Figure 5 As shown:
[0130] (1) (2) Shortest path at the portal layer, (3) Its total rough estimated cost is defined as:
[0131] ;
[0132] in, In the picture Above the border rights Calculate the shortest path;
[0133] S44: Candidate path retention; Since the rough estimation stage may be overly optimistic about unknown areas, a single route will not be directly adopted. To enhance reliability, the K candidate coverage paths with the lowest cost obtained in the rough estimation stage are selected. Reserved for subsequent actuarial calculations.
[0134] S45: Coverage path calculation stage;
[0135] For each candidate coverage path, perform an A* search on the real voxel map to calculate the real path cost:
[0136] ;
[0137] in, This represents the A* path distance calculated in free space. This represents the A* path distance calculated in the unknown space. As a penalty factor; To estimate feasible segmented paths, if a candidate path is infeasible, then let... Ultimately, the best option will be selected.
[0138] ;
[0139] S46: Global coverage path determination;
[0140] If a path includes sections that traverse unknown areas but are actually blocked by obstacles, its actual cost increases significantly or it may even become infeasible. A more accurate journey time can be obtained during the actuarial phase by running A* path planning once for each segment of the path, and then comparing the results. The precise calculation results of each path are used to select the path with the lowest cost as the final global coverage guide route. The precise calculation stage ensures the reliability of the global decision and prevents the path from being misled by rough estimates and ending up on a clearly suboptimal route.
[0141] The two-stage process described above is executed in each planning iteration, giving the global planning dynamic adaptive capabilities. When environmental information updates are minor, the coarse estimation stage can quickly update existing paths without recalculation, saving significant time. When major topology changes occur, the fine calculation stage provides leeway, not limiting itself to old paths, thus allowing for timely adjustments to the global strategy.
[0142] S5: Local exploration viewpoint planning based on global guidance, specifically including:
[0143] S51: Current target subspace determination: Determine the target subspace to be accessed based on the global coverage path order.
[0144] S52: Frontier Clustering and Candidate Viewpoint Sampling;
[0145] Within the target subspace, the leading edge region is clustered, and within each leading edge cluster, the space is sampled based on the sensor field of view to generate several candidate viewpoints.
[0146] S53: Calculation of viewpoint evaluation indicators;
[0147] The global coverage path provides global guidance for the local exploration planner, which optimizes the frontier access order and generates executable trajectories. Since high-quality viewpoints allow the UAV to update more environmental information (accessing more frontiers) and reduce exploration time, the frontier access order is determined by the viewpoint score. Therefore, viewpoint optimization is one of the key decisions for autonomous exploration. Compared to traditional evaluation methods based solely on information entropy or distance, the adaptive viewpoint evaluator model better balances the relationship between exploring new information and efficient advancement.
[0148] like Figure 6 As shown, for candidate viewpoints Perform a frustum projection and use ray projection to count the number of voxels falling into the unknown state within the field of view, denoted as . The viewpoint entropy gain can be defined as:
[0149] ;
[0150] The larger the value, the more new environmental information can be obtained from that viewpoint, and therefore the more beneficial it is for exploration;
[0151] like Figure 6 As shown, let the camera's field of view be FOV, and select K sampling direction vectors uniformly within the field of view. Ray projection is detected in all directions. If a ray preferentially contacts an obstacle, that direction is blocked; if it contacts the leading edge, it is considered a valid coverage ray. The visibility of these viewpoints is calculated:
[0152] ;
[0153] in, For finite coverage rays, K represents the total number of rays in the sampling direction. Viewpoints with high visibility indicate greater potential for observation of unknown space, which can improve exploration efficiency. Therefore, the following approach is introduced: It helps to reward those who have a broad and comprehensive perspective.
[0154] In this embodiment, distance cost Indicates the selection of viewpoint The required flight costs, based on the current drone location To viewpoint It is measured by the shortest path distance. Distance cost reflects the cost incurred: even if a viewpoint has a high information gain, if it is far away, the cost of traveling there immediately will be relatively high.
[0155] Define the current drone position as Candidate viewpoints are The shortest path distance between the two is:
[0156] ;
[0157] in, Indicates from arrive The set of all feasible paths, Representing an edge The weighted length takes into account factors such as flight costs, obstacle detours, and unknown penalties.
[0158] Normalize the distance:
[0159] ;
[0160] in: and These are the shortest and longest path distances in the candidate viewpoint set, respectively. It is a micro-term that prevents division by zero.
[0161] With negative weight in the scoring function Introducing a distance penalty is equivalent to penalizing distant viewpoints and encouraging nearby viewing points. This can be achieved through proper configuration. This allows for a balance between exploring new information and the cost of the journey, avoiding ineffective long-distance flights due to an excessive pursuit of information gain.
[0162] In this embodiment, small cluster penalty This section addresses the long-tail problem during exploration. The long tail refers to the scattered, uncovered areas that often remain near the end of an exploration, requiring drones to fly back and investigate, thus reducing efficiency. These scattered unknowns typically exist in small clusters, i.e., isolated and very small patches of unknown voxels. The aim is to reduce the priority of viewpoints targeting such small clusters. In practice, the total number of unknown voxels is calculated for each leading cluster. If the cluster size is below the threshold Then all viewpoints within that cluster are marked as "small cluster viewpoints" and assigned a fixed penalty increment. .
[0163] Let a certain front cluster be The number of its unknown voxels is If the cluster size is less than the threshold If so, it is determined to be the frontier of a small cluster:
[0164] ;
[0165] The small cluster penalty term is defined as:
[0166] ;
[0167] Small cluster penalty items The introduction of this feature allows the planner to tend to postpone processing these scattered areas, prioritizing the exploration of large, uncharted spaces. This prevents detours for areas with minimal gains and maintains the stability of the exploration process. However, small clusters are not permanently ignored; this embodiment incorporates a dedicated corner cleaning scheduler to clean these areas at appropriate times.
[0168] S54: Adaptive Viewpoint Score Calculation: Based on a weighted linear model, the above indicators are comprehensively scored to obtain candidate viewpoints. Final score The targets are sorted by score from highest to lowest to generate a local viewpoint target sequence. The final score is expressed as:
[0169] ;
[0170] All weights are satisfied:
[0171] ;
[0172] in, , , , These are the weighting parameters used to weigh the relative importance of information gain, view quality, distance cost, and small-scale cluster penalty;
[0173] S6: Local corner cleaning scheduling;
[0174] After obtaining an ordered list of viewpoint targets, the UAV also needs to consider cleaning up small corners when planning its local path. Near the end of autonomous exploration, a situation often arises where most of the space has been covered, but a few unexplored areas remain in certain local details (such as behind corners or in small cubicles). Conventional front-end driven methods might leave these scattered fronts until the end, processing them one by one when the main loop determines there are no large unknown areas to continue. This forces the UAV to take a long detour back to the previous area to finish, forming a long tail of exploration. This long tail directly reduces exploration efficiency and completion rate. Therefore, this embodiment designs an in-situ small corner cleaning scheduler as an enhancement component of the adaptive viewpoint evaluation model. It aims to balance viewpoint efficiency and propulsion stability, intelligently deciding when to clean up small corners in-situ. The scheduler inserts some low-value but necessary small fronts into the execution plan, cleaning them up in-situ without significantly affecting overall propulsion, avoiding long-distance backtracking later.
[0175] S61: Cleaning Trigger Detection: When the drone reaches the target viewpoint and completes the environmental scan, the cleaning scheduler is triggered;
[0176] S62: Neighborhood Small Cluster Front Detection: Within a preset neighborhood centered on the current location of the UAV, detect whether there is a front region that is identified as a small cluster;
[0177] S63: Nearest insertion sweep: If there is a small cluster front and no significant detour is required, the corresponding viewpoint is inserted into the current local viewpoint sequence, and sweep is performed with priority or second priority.
[0178] As shown in Figure 7(a), when the UAV reaches a new viewpoint and completes the environmental scan, a scheduler check is triggered. Within the current UAV's neighborhood R, a search is conducted to determine if a small cluster front exists. If it does, the scheduler directly adds the small cluster viewpoint to the second-highest scoring viewpoint in the local planner's viewpoint list (below the viewpoint in the next step). The UAV then flies towards this small front for observation coverage and continues to the next target. Because the range of R is finite, the additional cost of this deviation is small, but it can effectively eliminate potential long tails in a timely manner.
[0179] S64: Orientation Cleaning: If the leading edge of the small cluster can be covered by adjusting the drone's orientation in place, a cleaning action containing only attitude changes is generated to complete the cleaning without changing the flight position.
[0180] like Figure 7 As shown in (b), the drone reaches a corridor corner and prepares to proceed along the corridor, but there may be a small section of wall behind it that is not clearly visible, meaning there is a small cluster of fronts in the opposite direction of the current viewpoint. The scheduler adds this reverse viewpoint to the highest-scoring viewpoint in the local planner's viewpoint list, and the drone turns around at a certain angle in place to clear that corner. This avoids revisiting the area later because of this oversight. Orientation cleaning has a very low cost (rotating in place), making it a very effective method to improve coverage efficiency.
[0181] The scheduler follows a conditional triggering mechanism, executing cleaning only when the detected small leading edge matches a cluster leading edge and is within the cleaning area R. If the small leading edge is far away or requires a significant detour, the scheduler chooses to preserve it, allowing the main planning to continue (in this case, these leading edges are still...). (With penalties excluded), this design ensures the steady progress of the main exploration storyline while eliminating long-tail problems that could seriously slow down efficiency in the early stages.
[0182] S7: Trajectory planning, execution, and environment updates;
[0183] S71: Path search: Based on the current local viewpoint sequence, a voxel-level A* search is used for near targets, and a graph A* search is used for far targets to generate a navigation path;
[0184] S72: Trajectory Generation: Based on the generated path, construct a B-spline trajectory that satisfies the constraints of smoothness, safety, and UAV dynamics.
[0185] S73: Flight execution and map updates;
[0186] The drone flies along the trajectory while updating the voxel map, frontier information, and connectivity structure in real time.
[0187] S74: Planning Cycle:
[0188] If environmental information changes or the current global coverage path fails, return to perform spatial decomposition and planning again;
[0189] Repeat the above steps until there are no more unexplored frontiers, at which point the exploration mission is complete.
[0190] This embodiment provides spatial topology layering information (such as connectivity graphs and sub-region divisions) and a candidate front viewpoint list, which are the basic inputs for global and local planning. The order of target areas output by the global coverage path planning directly affects the selection and sorting of local viewpoints. Local planning decisions, in turn, affect the execution and adjustment of the global path (for example, if the local planning deviates too much from the global guidance, the global path needs to be reconstructed). In addition, the introduced corner cleaning scheduler runs through the decision-making process, checking whether there are any remaining uncovered small areas after each target viewpoint is reached, and inserting additional cleaning actions as appropriate.
[0191] Example 2
[0192] This embodiment provides a layered autonomous exploration system for unmanned aerial vehicles (UAVs) based on coverage path guidance, which is used to implement the layered autonomous exploration method for UAVs based on coverage path guidance in Embodiment 1 above. It includes: a data acquisition module, a modeling module, a portal hierarchy graph construction module, a portal layer connectivity graph construction module, a global coverage path planning module, a local exploration viewpoint planning module, a cleaning scheduling module, and an exploration module.
[0193] In this embodiment, the data acquisition module is used to acquire UAV pose information and environmental observation data;
[0194] In this embodiment, the modeling module is used to construct an environmental perception and three-dimensional voxel map;
[0195] In this embodiment, the portal hierarchy graph construction module is used to construct a portal hierarchy graph, where each node represents a connected subspace, and each edge represents a passable connection between subspaces established through the portal, generating free regions and unknown regions.
[0196] In this embodiment, the portal layer connectivity graph construction module is used to construct the portal layer connectivity graph by using the centers of free regions and unknown regions as nodes of the portal layer graph;
[0197] In this embodiment, the global coverage path planning module is used to perform global coverage path planning based on the portal layer connectivity graph;
[0198] In this embodiment, the local exploration viewpoint planning module is used to determine the current target subspace based on the global coverage path order, generate a local exploration viewpoint target sequence, and perform local exploration viewpoint planning;
[0199] In this embodiment, the cleaning scheduling module is used to schedule local corner cleaning;
[0200] In this embodiment, the exploration module is used to search for targets based on the current local exploration viewpoint target sequence, generate a navigation path, and update the three-dimensional voxel map.
[0201] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.
Claims
1. A hierarchical autonomous exploration method for unmanned aerial vehicles (UAVs) based on coverage path guidance, characterized in that, Includes the following steps: Acquire UAV pose information and environmental observation data to construct environmental perception and 3D voxel maps; Construct a portal hierarchy graph, where each node represents a connected subspace and each edge represents a passable connection between subspaces established through the portal, generating free and unknown regions; By using the centers of free and unknown regions as nodes in the portal layer graph, a portal layer connectivity graph is constructed. Global coverage path planning based on portal layer connectivity graph; Based on the global coverage path order, the current target subspace is determined, a local exploration viewpoint target sequence is generated, and local exploration viewpoint planning is performed. Conduct on-site corner cleaning and dispatch; The target is searched based on the current local exploration viewpoint target sequence, a navigation path is generated, and the 3D voxel map is updated.
2. The hierarchical autonomous exploration method for unmanned aerial vehicles based on coverage path guidance according to claim 1, characterized in that, Acquire UAV pose information and environmental observation data to construct environmental perception and 3D voxel maps, specifically including: Environmental observation data is acquired using airborne visual sensors on UAVs, and the perception range is modeled as a three-dimensional view frustum. The exploration space is represented in voxel form using an occupied grid map, which is divided into multiple three-dimensional voxel units. The voxel states are divided into free voxels, occupied voxels, and unknown voxels. The boundary region between free voxels and unknown voxels is taken as the frontier region.
3. The hierarchical autonomous exploration method for unmanned aerial vehicles based on coverage path guidance according to claim 1, characterized in that, Constructing a portal hierarchy diagram specifically includes: The space to be explored is divided into multiple unit spaces. When it is determined that the 3D voxel map has been updated, the unit spaces are decomposed. Within the decomposed cell space, perform three-dimensional connected component labeling on voxels; Based on the labeling results of the three-dimensional connected components, voxels with the same label are aggregated to form regions, regions composed of safe and free voxels are defined as free regions, and regions composed of unknown voxels are defined as unknown regions. For each region, the corresponding voxel geometric center is calculated as the initial region center, and the initial region center is corrected.
4. The hierarchical autonomous exploration method for unmanned aerial vehicles based on coverage path guidance according to claim 1, characterized in that, Constructing the portal layer connectivity graph specifically includes: Use the center of the free region as the vertex and the connectivity between free regions as the edge; Construct connected edges for the unknown region; For a spatial cell that contains both free and unknown regions, perform a restricted A* search within the cell between the corresponding free and unknown nodes. If the cell is determined to be reachable, establish a portal edge.
5. The hierarchical autonomous exploration method for unmanned aerial vehicles based on coverage path guidance according to claim 1, characterized in that, Global coverage path planning based on the portal layer connectivity graph specifically includes: Determine the set of target nodes for the coverage path, and use the incremental A* algorithm to search for the path on the portal layer connectivity graph to obtain the coverage path; Calculate the cost of the coverage path and retain the K candidate coverage paths with the minimum cost; For each candidate coverage path, perform an A* search on the voxel map to calculate the actual path cost; Compare The path with the lowest cost is selected from the calculated results of each path as the final global coverage path guide route.
6. The hierarchical autonomous exploration method for unmanned aerial vehicles based on coverage path guidance according to claim 5, characterized in that, The cost of the covered path is calculated as follows: ； in, In the picture The shortest path calculated above, Indicates the current location of the drone. Indicates the center of the candidate viewpoint. This represents the edge-to-edge cost given by the path length A*.
7. The hierarchical autonomous exploration method for unmanned aerial vehicles based on coverage path guidance according to claim 5, characterized in that, Perform an A* search on the voxel map to calculate the true path cost, specifically as follows: ； in, This represents the A* path distance calculated in free space. This represents the A* path distance calculated in the unknown space. As a penalty factor, To estimate the feasible segmented path, Indicates the current location of the drone. This represents K candidate coverage paths.
8. The hierarchical autonomous exploration method for unmanned aerial vehicles based on coverage path guidance according to claim 1, characterized in that, Based on the global coverage path order, the current target subspace is determined, and local exploration viewpoint planning is performed, specifically including: Within the target subspace, the leading edge region is clustered, and within each leading edge cluster, the space is sampled according to the sensor field of view to generate several candidate viewpoints; Calculate viewpoint evaluation metrics and adaptive viewpoint scores; The adaptive viewpoint scores are sorted to obtain a local exploration viewpoint target sequence, and local exploration viewpoint planning is performed based on the local exploration viewpoint target sequence.
9. The hierarchical autonomous exploration method for unmanned aerial vehicles based on coverage path guidance according to claim 7, characterized in that, Calculate the viewpoint evaluation index and the adaptive viewpoint score, expressed as: ； ； ； ； ； ； ； in, , , , For each weight parameter, Represents viewpoint entropy gain. Indicates the number of voxels. For finite coverage rays, K is the total number of rays in the sampling direction. Indicates the visibility of the viewpoint. Represents the normalized distance. Indicates the current location of the drone. The shortest path distance to candidate viewpoint v Indicates from arrive The set of all feasible paths, Representing an edge The weighted length, It is a minor term to prevent division by zero. and These are the shortest and longest path distances in the candidate viewpoint set, respectively. This represents the small cluster penalty term. Indicates a frontier cluster, Indicates the number of unknown voxels. This represents the cluster size threshold.
10. A hierarchical autonomous exploration system for unmanned aerial vehicles based on covered path guidance, characterized in that, The method for implementing the UAV hierarchical autonomous exploration method based on coverage path guidance as described in any one of claims 1-9 includes: a data acquisition module, a modeling module, a portal hierarchy graph construction module, a portal layer connectivity graph construction module, a global coverage path planning module, a local exploration viewpoint planning module, a cleaning scheduling module, and an exploration module. The data acquisition module is used to acquire UAV pose information and environmental observation data; The modeling module is used to construct environmental perception and three-dimensional voxel maps; The portal hierarchy graph construction module is used to construct a portal hierarchy graph, where each node represents a connected subspace, each edge represents a passable connection between subspaces established through the portal, and free regions and unknown regions are generated. The portal layer connectivity graph construction module is used to construct the portal layer connectivity graph by using the free region center and the unknown region center as nodes of the portal layer graph. The global coverage path planning module is used to perform global coverage path planning based on the portal layer connectivity graph. The local exploration viewpoint planning module is used to determine the current target subspace based on the global coverage path order, generate a local exploration viewpoint target sequence, and perform local exploration viewpoint planning. The cleaning scheduling module is used to schedule local corner cleaning. The exploration module is used to search for targets based on the current local exploration viewpoint target sequence, generate navigation paths, and update the three-dimensional voxel map.