An unmanned aerial vehicle adaptive autonomous exploration method based on local view topology optimization

Abstract

The application discloses an unmanned aerial vehicle self-adaptive autonomous exploration method based on local view topology optimization, relates to the technical field of unmanned aerial vehicle autonomous environment perception and robot path planning, and synchronously collects multi-source sensing data to construct TSDF and ESDF voxel maps; free space regions are divided and passable portals are extracted through state machine incremental topology decomposition; a hybrid granularity global access sequence is constructed by dividing far and near regions according to the local view; the viewpoint planning cost is adaptively adjusted in combination with front information gain; a cubic B-spline smooth trajectory is generated according to dynamic constraints and is cyclically executed until environment exploration is completed. The incremental topology structure reduces repeated calculation, the portal node optimizes the flight path, the adaptive viewpoint planning avoids local wandering, lightweight solution adapts to onboard equipment, and the exploration time length and flight distance are shortened while complete environment coverage is ensured.

Description

Technical Field

[0001] This invention relates to the field of autonomous environmental perception and robot path planning technology for unmanned aerial vehicles (UAVs), specifically to an adaptive autonomous exploration method for UAVs based on local horizon topology optimization. Background Technology

[0002] Small unmanned aerial vehicles (UAVs) equipped with depth sensing devices can autonomously explore enclosed, unknown environments without GPS signals, and are widely used in disaster relief, underground mine inspection, and indoor industrial equipment inspection. The complete autonomous exploration process for UAVs includes six modules: multi-source sensor data acquisition, 3D voxel map reconstruction, free space region division, global access path planning, local observation viewpoint ranking, smooth trajectory generation, and flight control. The device relies on depth cameras, odometry, and inertial measurement units to collect environmental geometry and its own pose information. It uses TSDF voxel meshes to characterize the environment, distinguishing between unknown, free, and occupied spatial states. It extracts the leading edge clusters at the boundary between free and unknown spaces as subsequent exploration targets, and uses a combined optimization algorithm to output the UAV's flight path, achieving full environmental coverage detection. While existing mainstream exploration frameworks can stably complete environmental traversal in open, single spaces, their stability and spatial exploration efficiency are significantly limited when facing complex maze scenarios with multiple room partitions, narrow corridors, non-convex walls, and multiple intersecting intersections.

[0003] Existing solutions generally maintain free space connectivity by globally reconstructing the topology map. After each map perception update, the system clears all historical spatial partitioning results and re-executes the operations of 3D connected component labeling, region center point extraction, and adjacent region connected path search. As the drone continues to fly, the environmental voxel state frequently undergoes incremental changes, causing the global reconstruction mechanism to continuously generate high computational overhead and form periodic computing power peaks. At the same time, repeated re-clustering causes slight drift in the position of region center points, and the connectivity of adjacent regions changes synchronously. The access sequence output by global path planning continues to oscillate, and the flight trajectory is frequently adjusted, increasing invalid round-trip flight behavior. Traditional topology simplification methods only use a single center point to represent the entire free space region, ignoring real passable geometric structures such as doorways and corridors. The lines between center points do not represent actual flyable channels, and the access sequence output by macro-path planning is prone to detours and backtracking problems during landing execution.

[0004] Traditional local viewpoint ranking relies solely on flight distance, flight duration, and changes in fuselage yaw angle to construct a cost matrix. This algorithm inherently favors visiting the nearest forward target, easily falling into a locally greedy state at intersections and room corners, neglecting larger, more valuable, and unknown areas in the distance. At the global planning level, there's a trade-off between modeling accuracy and real-time solution speed. If all region entrance and exit nodes are fully retained for combinatorial optimization, the problem size rapidly expands, exceeding the real-time computational demands of airborne embedded devices. Conversely, if only region center points are included, path feasibility significantly decreases. Existing technologies lack a hierarchical, granular adaptive solution mechanism, making it difficult to simultaneously balance near-field path accuracy and long-distance planning computational efficiency. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive autonomous exploration method for unmanned aerial vehicles (UAVs) based on local horizon topology optimization, so as to solve the problems mentioned in the background art.

[0006] An adaptive autonomous exploration method for unmanned aerial vehicles (UAVs) based on local horizon topology optimization includes the following complete execution steps: S1 performs time registration of depth images, odometry pose data and IMU (Inertial Measurement Unit) inertial data, constructs TSDF (Truncation Signed Distance Function) voxel mesh and incrementally solves ESDF (Euclidean Signed Distance Field) distance field, divides voxel occupancy states, and stores voxel connectivity labels, truncation distances, cumulative weights and passage distance constraints. S2 filters voxels with state changes within the bounding box, extracts free space connected components through 3D connected component labeling and obtains the center point of Zone (topology partition), performs threshold matching between the current center point and the historical center point, distinguishes the three types of Zone states and performs corresponding topology update operations, and removes tiny connected components with too small an area. S3 constructs connecting edges between adjacent Zones, and solves for traversable paths using local A*, incremental edge computation, or equivalent collision-free path search methods. The midpoint of each path is extracted as a representative Portal point and recorded in the Portal set of adjacent Zone nodes. During Portal generation, path search collision constraints, free space filtering, and ESDF / occupied map constraints are combined to ensure the traversability of connecting paths. Zones without valid Portal records use their own center point as planning nodes. S4 defines the local horizon based on the drone's location, dividing the Zone into near and far zones. The near zone uses Portal nodes, while the far zone only retains the center point. If the number of Portals does not exceed the limit, a Portal-GTSP (Gradient Time Synchronization Protocol) model is constructed and converted into an ATSP (Asymmetric Traveling Salesman Problem) model to solve the path. If the limit is exceeded, the Portals are simplified and degenerated into center point planning, outputting a hybrid access sequence. S5 extracts the environmental frontier clusters and calculates the information gain. After normalization, it constructs a cost matrix by combining flight distance and attitude cost. It prioritizes high information frontier viewpoints, eliminates candidate points that violate flight constraints, and solves the local viewpoint access order. S6 combines the viewpoint sequence with UAV dynamic constraints to construct a cubic B-spline smooth trajectory and optimize the yaw angle curve. It periodically issues flight control commands and iteratively executes S1 to S6 until the space exploration coverage is achieved.

[0007] Furthermore, in step S1, the TSDF voxel update uses a weighted iterative pruning calculation, and the calculation formula is as follows: ; In the formula This represents the TSDF value of the (i)th voxel at time (t). This represents the TSDF value of the voxel at the previous time step. Indicates the historical cumulative weight. This indicates the update weight corresponding to the current depth observation. This represents the voxel truncation symbol distance obtained from the depth observation of the current frame. Indicates the TSDF cutoff distance. This indicates that the calculation results will be restricted to... Within the range.

[0008] Furthermore, the core calculation formula for the frontier information gain normalization and reconstruction cost matrix in step S5 is as follows: In the formula, Representing the Normalized information gain of frontier clusters, Indicates the first A frontier cluster, This indicates the number of voxels contained in the frontier cluster. Indicates the first Information gain of a frontier cluster This represents the maximum information gain among the candidate frontier clusters currently participating in SOP sorting. To avoid extremely small constants with a denominator of zero, Indicates viewpoint Flying to the frontiers or candidate viewpoints The original physical cost can be composed of factors such as flight distance, flight time, and yaw variation; This represents the reconstruction cost after information gain discounting; This represents the dynamic greedy coefficient for the current planning cycle; Indicates the frontier of the target The normalized information gain. The higher the information gain at the target frontier, the better. The smaller the value, the more likely it is to be prioritized in the SOP sorting process.

[0009] Furthermore, in step S2, the matching of the old and new center points of the Zone uses the Euclidean distance determination rule, and a fixed matching distance threshold is set. For any Zone center point in the current period, traverse all historical Zone center points from the previous period. The Euclidean distance between any two points is less than 1. If a zone is determined to be an evolution of the same connected region, it directly inherits the connected edges, Portal set, and unknown voxel statistics of the historical Zone. If a single Zone center point matches multiple historical Zone center points in the current period, it is determined to be a region merging state. Only the connected edges of the Zones with association are locally recalculated. If the current center point has no matching historical nodes, it is determined to be a new Zone. A new topology node is assigned and the search for surrounding connected edges is started.

[0010] Furthermore, in step S3, the Portal representative point extraction process is completed synchronously with the Zone connected edges. When two Zones establish a connection, a local A* path search is initiated. After the path traversal is completed, the middle position of the path coordinate sequence is taken as the Portal representative point and stored in the built-in Portal set of the two Zone nodes respectively. In the subsequent global GTSP planning stage, the Portal set is directly read as the candidate node within the cluster. For Zones that have not generated a Portal, the region center point is used as the fallback node. The Portal stores the real geometric position of the passageway, making up for the lack of path reachability caused by relying solely on the abstract space of the Zone center point.

[0011] Furthermore, step S4 divides the region into near-zone and far-zone based on the horizon radius. The near-zone uses all Portals for GTSP calculation, while the far-zone uses only the zone center point for ATSP calculation. A threshold for the number of Portals is set. If the threshold is exceeded, representative Portals are selected based on indicators such as the distance between the Portal and the UAV, the number of unknown voxels in the zone, and the cost of connected paths. If the number still exceeds the threshold after selection, the process is switched to center point planning to control the problem size. The GTSP model is converted to an ATSP model using the Noon-Bean transformation, and the LKH solver is called to obtain the traversal order of the zones and Portals. This macro-sequence only limits the traversal order, while the flight trajectory is refined by the local viewpoint planning module. Layered planning reduces the solution dimension.

[0012] Furthermore, in step S5, the dynamic greed coefficient is adaptively adjusted according to the number of maximum voxels at the local front, and the dynamic greed coefficient is calculated using a saturation function. In the formula This represents the information greed coefficient for the current planning period. This represents the maximum greed coefficient. This represents the maximum number of front voxels in the current set of candidate fronts participating in SOP sorting. This represents the saturation adjustment constant. When the current local candidate front is large... As the size of the local front increases, the system tends to prioritize accessing regions with high information gain; when the size of the local front decreases... Automatically reduce speed, and planning gradually returns to prioritizing distance; Step S5 sets up multi-layer physical constraints to filter candidate viewpoints. Candidate viewpoints must simultaneously satisfy the following conditions: ESDF distance is greater than the safe passage distance, the difference in yaw angle between adjacent viewpoints is less than the maximum yaw change, and the leading edge region is within the camera's field of view. If any constraint is not met, the corresponding transfer cost is set to infinity, and the viewpoint will no longer participate in the SOP sorting calculation, thus eliminating invalid viewpoints in advance.

[0013] Furthermore, step S6 uses cubic B-splines to complete trajectory parameterization, uniformly sets fixed control point spacing, and simultaneously applies four-dimensional dynamic constraints of maximum linear velocity, maximum linear acceleration, maximum yaw rate, and maximum yaw rate during the trajectory optimization stage. The position trajectory and yaw trajectory are optimized separately to generate smooth curves. If the map update causes a significant change in the Zone topology or a significant change in the total number of leading clusters during a single trajectory execution, the current trajectory is immediately terminated and global path planning and local viewpoint sorting are re-executed to dynamically respond to spatial structure changes caused by environmental perception.

[0014] Furthermore, in step S2, after the three-dimensional connected components are marked, a small connected domain filtering mechanism is added. When the total number of free voxels contained in a connected domain is less than the minimum voxel threshold, it is not registered as a valid Zone node. At the same time, the unknown connected domain filtering threshold is calculated based on the voxel grid size, CCL sampling step size, and minimum unknown scale coefficient to remove scattered unknown regions.

[0015] Furthermore, this method does not rely on GPS signals throughout the entire process, but only on an airborne depth camera, odometry and IMU to achieve perception and positioning. The algorithm is deployed on a small quadcopter UAV embedded platform. The voxel map adopts a dual-layer structure with a fine resolution of 0.1m and a coarse resolution of 0.2m. The observation distance, TSDF cutoff distance, grid size and sampling step size are all adaptively adjusted according to the scene and computing power.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention introduces a regional state machine to complete incremental topology maintenance, performs local processing on voxels whose states change after map updates, and divides the spatial change states into three types: expansion, merging, and new creation by matching the distance between the center points of the old and new regions. It preserves the historical topology structure and connectivity, and only performs local connectivity edge repair work on spatial merging and newly added regions. It does not require global reconstruction of the topology map after each map update, which can reduce the system's computational load, eliminate periodic computing power peaks, and avoid path sequence fluctuations caused by the drift of regional center points, thereby improving the long-term stability of the topology structure in complex environments.

[0017] This invention extracts representative points of passable channels as Portal nodes when establishing connectivity between adjacent free regions. It expands the abstract region of a single center point into a composite spatial representation structure with a center point and multiple Portals. During the global path planning stage, Portal nodes can be selected as candidate points for access within the region, restoring the real passable geometric features of the space. This alleviates the detours and backtracking caused by path planning in non-convex rooms and narrow corridors, and improves the matching degree between the global access sequence and the actual flight environment.

[0018] This invention sets up a local horizon to divide the area into near and far regions and performs hybrid granularity topology planning. The near-field area to be visited by the UAV retains all Portal nodes to participate in high-precision combinatorial optimization, while the far-field area only relies on the center point of the region to provide macroscopic access direction. At the same time, a threshold for the number of Portals is set to dynamically adjust the solution model. When the number of Portals in the near area exceeds the threshold, representative Portals are selected or the solution is switched to the center point. This maintains the physical access accuracy of the near path and controls the scale of the combinatorial optimization problem, adapting to the real-time computing constraints of the airborne embedded platform.

[0019] This invention constructs an information-driven local viewpoint ranking logic, which measures the information gain of a region by the total number of voxels contained in the frontier cluster, and adaptively adjusts the information weight based on the overall scale of the current local frontier. In the stage of concentrated unknown space, high-information regions are prioritized for access. As the size of the unknown region decreases, the logic gradually switches to distance-first planning, which alleviates the detection omission problem caused by local greed in traditional algorithms and takes into account the needs of large-scale unknown space search in the early stage of exploration and path convergence optimization in the later stage.

[0020] This invention applies multiple physical constraints to candidate observation viewpoints, eliminating invalid viewpoints in advance that pose a collision risk, cannot be observed by the camera in unknown areas, or whose fuselage orientation does not meet motion restrictions. This reduces meaningless computations in the local sorting stage. The trajectory generation module uses cubic B-splines to independently optimize the position and yaw angle curves, strictly matching the UAV dynamic boundary conditions to generate smooth, safe, and directly executable flight trajectories. When the map topology changes significantly, the planning process can be restarted immediately, dynamically adapting to spatial structure changes brought about by environmental perception, and optimizing overall exploration efficiency while maintaining complete environmental coverage. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the overall process of adaptive autonomous exploration by unmanned aerial vehicles based on topology optimization proposed in this invention. Figure 2 This is a flowchart of the incremental spatial topology decomposition and portal extraction process of the present invention; Figure 3 This is a flowchart of the hybrid granularity global planning and adaptive local viewpoint planning of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figure 1-3 This invention provides a technical solution: an adaptive autonomous exploration method for unmanned aerial vehicles (UAVs) based on local horizon topology optimization, comprising the following complete execution steps: Step S1: Multi-source sensor data synchronous input and voxel map initialization; acquire depth images output by depth camera, pose data from airborne odometry, and inertial measurement data from IMU and complete time alignment; construct a TSDF truncated symbolic distance voxel grid based on ray projection; solve the ESDF Euclidean distance field incrementally from TSDF; divide all three-dimensional voxels into three states: unknown, free, and occupied; synchronously record the Zone connected region label, truncation distance, cumulative weight, and safe passage distance constraint corresponding to each voxel. Step S2 implements incremental spatial topology decomposition based on the Zone state machine. It filters voxels whose states are flipped within the map update bounding box, uses a three-dimensional connected component labeling algorithm to extract free space connected domains to obtain the current Zone center point set, and compares the current center points with the historical center points of the previous cycle one by one according to the matching distance threshold. Based on the matching results, it divides the Zone state into three categories: expansion and stability, region merging, and newly generated. Correspondingly, it performs historical topology anchor point inheritance, local connected edge repair, and new topology node registration operations, filters out small connected domains with insufficient voxels, retains historical topology memory throughout the process, and only recalculates local connected relationships affected by map updates. Step S3: Portal extraction and drivability verification of adjacent Zone nodes. When establishing a connecting edge between any two Zones, the local A*, incremental edge calculation, or equivalent collision-free path search method is called to solve the collision-free drivable path between the two regions. The midpoint of the path is extracted as the representative point of the Portal and recorded in the Portal set of adjacent Zone nodes. During the Portal generation process, the drivability of the connecting path is guaranteed by combining path search collision constraints, free space filtering, and ESDF / occupied map constraints. Zones without valid Portals default to using their own center point as global planning candidate nodes. Step S4 involves local horizon constraint hybrid granularity topology global path planning. A fixed local horizon radius is set based on the real-time position of the UAV. The spatial zone is divided into a near zone within the horizon and a far zone outside the horizon. All Portals in the near zone are extracted as planning candidate nodes, while only the zone center point is retained as a low-dimensional macroscopic guiding node in the far zone. A maximum threshold for the number of Portals in the near zone is set. If the total number of Portals does not exceed the limit, a Portal-GTSP generalized traveling salesman model is directly constructed. After Noon-Bean transformation, it is converted into an ATSP model and then the LKH heuristic solver is called to calculate the access sequence. If the total number of Portals exceeds the threshold, representative Portals are selected or the model is reverted to using only the center point of the ATSP model. The global zone and Portal hybrid access sequence is output. Step S5 is an information-driven adaptive local viewpoint ranking planning process. Under the constraints of the global access sequence, all frontier clusters in the map are extracted. The total number of voxels contained in a single cluster is used as the original information gain. Local normalization is performed on all candidate frontiers in the current period to eliminate scale differences. The dynamic greedy coefficient is calculated based on the number of voxels in the local maximum frontier. Combined with the changes in flight distance and yaw angle, the original physical transfer cost is constructed. The information gain discount is introduced to reconstruct the SOP sequence ranking cost matrix, which improves the access priority of high information frontiers. Invalid candidate viewpoints that do not meet the constraints of camera field of view, safe distance, and yaw angle velocity are filtered out. The short-term local viewpoint access order is obtained by solving the problem. Step S6 involves the generation and execution of a cubic B-spline smoothed trajectory under dynamic constraints. A cubic B-spline position trajectory is constructed based on the local viewpoint sequence, the UAV velocity and acceleration boundaries, and the safety distance limit. The yaw angle smoothing curve is optimized separately. After optimization, position, velocity, acceleration, and yaw control commands are sent to the flight control system at fixed intervals. After a single short-time trajectory is executed, steps S1 to S6 are executed repeatedly until the total amount of the unknown area front edge of the map is lower than the preset threshold or the spatial coverage volume reaches the target standard.

[0024] In this invention, the TSDF voxel update in step S1 uses a weighted iterative pruning calculation, and the calculation formula is as follows: ; In the formula This represents the TSDF value of the (i)th voxel at time (t). This represents the TSDF value of the voxel at the previous time step. Indicates the historical cumulative weight. This indicates the update weight corresponding to the current depth observation. This represents the voxel truncation symbol distance obtained from the depth observation of the current frame. Indicates the TSDF cutoff distance. This indicates that the calculation results will be restricted to... Within this range, the iterative update method can continuously fuse multiple frames of depth observations, reducing the interference of single-frame depth noise on the accuracy of voxel maps.

[0025] In this invention, the core calculation formula for the frontier information gain normalization and reconstruction cost matrix in step S5 is as follows: In the formula, Representing the Normalized information gain of frontier clusters, Indicates the first A frontier cluster, This indicates the number of voxels contained in the frontier cluster. Indicates the first Information gain of a frontier cluster This represents the maximum information gain among the candidate frontier clusters currently participating in SOP sorting. To avoid extremely small constants with a denominator of zero, Indicates viewpoint Flying to the frontiers or candidate viewpoints The original physical cost can be composed of factors such as flight distance, flight time, and yaw variation; This represents the reconstruction cost after information gain discounting; This represents the dynamic greedy coefficient for the current planning cycle; Indicates the frontier of the target The normalized information gain. The higher the information gain at the target frontier, the better. The smaller the value, the more likely it is to be prioritized in the SOP ranking. Normalized gain can unify the information weights of frontiers of different scales, and reconstructing the cost matrix can dynamically increase the access priority of regions with highly unknown information.

[0026] In this invention, the matching of the old and new center points of the Zone in step S2 adopts the Euclidean distance determination rule and sets a fixed matching distance threshold. For any Zone center point in the current period, traverse all historical Zone center points from the previous period. The Euclidean distance between any two points is less than 1. If a zone is determined to be evolving within the same connected region, it directly inherits the connected edges, Portal set, and unknown voxel statistics of the historical Zone. If a single Zone center point matches multiple historical Zone center points in the current period, it is determined to be a region merging state. Only the connected edges of the Zones with associations are locally recalculated, without globally clearing and rebuilding the topology. If the current center point has no matching historical nodes, it is determined to be a new Zone. A new topology node is assigned and a search for surrounding connected edges is started. This incremental topology maintenance mechanism eliminates the global computing power peak caused by high-frequency map updates and suppresses the path sequence oscillation caused by Zone center point drift.

[0027] In this invention, the Portal representative point extraction process in step S3 is completed synchronously with the Zone connected edges. When two Zones establish a connection, a local A* path search is initiated. After the path traversal is completed, the middle position of the path coordinate sequence is taken as the Portal representative point and stored in the built-in Portal set of the two Zone nodes respectively. In the subsequent global GTSP planning stage, the Portal set is directly read as the candidate node within the cluster. For Zones that have not generated a Portal, the region center point is used as the fallback node. The Portal stores the real geometric position of the passageway, which makes up for the lack of path reachability caused by relying solely on the abstract space of the Zone center point and reduces flight detours and backtracking behaviors in non-convex rooms and narrow corridors.

[0028] In this invention, step S4 involves dividing the local view radius into near and far zones and performing hybrid granularity planning. Near zones that are within the view radius of the UAV's current position fully utilize all internal Portals for GTSP solution, while far zones that are beyond the view radius only use the zone center point for ATSP solution. At the same time, a threshold is set for the number of near-zone Portals. When the total number of Portals participating in the calculation exceeds the threshold, some Portals are filtered based on the dual indicators of the unknown space volume of the zone and the distance between the Portal and the UAV. If the total number still exceeds the limit after filtering, the Portal nodes are abandoned, and the solution is uniformly switched to center point ATSP solution. The hybrid granularity strategy takes into account both the physical access accuracy of the near path and the computational efficiency of the long-distance planning, preventing the number of Portals from causing the combinatorial optimization problem to become out of control. The conversion from the GTSP model to the ATSP model uses the standard Noon-Bean transformation algorithm. After the transformation, the open-source LKH heuristic solver is called to solve for the optimal access sequence. The macroscopic access sequence output by the solver only constrains the traversal order of Zone and Portal, and does not directly generate the flight trajectory. Subsequently, the local viewpoint sorting module refines the leading observation points in a single region. The hierarchical planning architecture splits the global macroscopic path and local fine observation tasks, reduces the solution dimension of a single optimization problem, and balances the integrity of global exploration with the smoothness of local trajectory.

[0029] In this invention, the dynamic greed coefficient in step S5 is adaptively adjusted according to the number of voxels with the maximum local frontier, and the dynamic greed coefficient is calculated using a saturation function. In the formula This represents the information greed coefficient for the current planning period. This represents the maximum greed coefficient. This represents the maximum number of front voxels in the current set of candidate fronts participating in SOP sorting. This represents the saturation adjustment constant. When the current local candidate front is large... As the size of the local front increases, the system tends to prioritize accessing regions with high information gain; when the size of the local front decreases... Automatic descent and gradual return to distance priority in planning enable adaptive switching between actively searching for unknown spaces in the early stages of exploration and converging and optimizing flight paths in the later stages. Step S5 filters valid candidate viewpoints by setting multi-layered physical constraints. Any candidate viewpoint must simultaneously meet three conditions: the ESDF distance at the viewpoint is greater than the safe passage distance, the difference in yaw angle between the viewpoint and the previous viewpoint is less than the maximum allowable yaw change, and the effective part of the candidate leading edge falls within the field of view of the UAV camera and meets the minimum number of visible voxels requirement. If any constraint is not met, the transfer cost corresponding to that viewpoint is assigned to infinity and it will not participate in the SOP sorting solution. Multi-layered constraints eliminate invalid viewpoints with collision risks, unreasonable orientations, or inability to observe unknown areas in advance, reducing the amount of invalid computation in the local planning stage and ensuring that the generated trajectory can be executed safely.

[0030] In this invention, step S6 uses cubic B-splines to complete trajectory parameterization, uniformly sets fixed control point spacing, and simultaneously applies four-dimensional dynamic constraints of maximum linear velocity, maximum linear acceleration, maximum yaw rate, and maximum yaw rate during the trajectory optimization stage. The position trajectory and yaw trajectory are optimized separately to generate smooth curves. If the map update causes a significant change in the Zone topology or a significant change in the total number of leading clusters during a single trajectory execution, the current trajectory is immediately terminated and global path planning and local viewpoint sorting are re-executed. This dynamically responds to spatial structure changes caused by environmental perception, ensuring that the flight trajectory continuously matches the real-time environmental state.

[0031] In this invention, step S2 adds a small connected domain filtering mechanism after the three-dimensional connected components are marked. When the total number of free voxels in a connected domain is less than the minimum voxel threshold, it is not registered as a valid Zone node. At the same time, the unknown connected domain filtering threshold is calculated based on the voxel grid size, CCL sampling step size, and minimum unknown scale coefficient to eliminate scattered unknown regions that are too small and have no effective exploration value, thereby reducing the number of redundant nodes in the topology graph, further compressing the computation time of global GTSP and ATSP optimization solutions, and improving the real-time operation stability of the system in complex maze and multi-branch scenarios.

[0032] In this invention, the entire method does not require GPS positioning signals and relies solely on an airborne depth camera, odometer, and IMU to complete all perception and positioning calculations. The hardware can be mounted on a small quadcopter UAV embedded computing platform. The voxel map supports dual-layer representation with a fine resolution of 0.1m and a coarse resolution of 0.2m. The maximum observation distance of ray projection, TSDF truncation distance, maximum cell size of hierarchical grid, and CCL sampling step size can all be flexibly configured according to the scene size and the computing power of the UAV payload. It is suitable for various autonomous exploration tasks in closed unknown environments without GPS, such as disaster relief, underground mines, and indoor industrial inspection.

[0033] This invention presents an autonomous exploration method for unmanned aerial vehicles (UAVs) based on local horizon adaptive perception and topology optimization. The entire method relies on an onboard depth camera, odometry, and IMU to achieve GPS-free environmental perception and positioning. A multi-core embedded computing unit performs all computations, and the entire process is divided into six complete execution flows: sensor input mapping, incremental topology decomposition, portal extraction, hybrid granularity global planning, information-driven local viewpoint ranking, cubic B-spline trajectory generation, and iterative iteration. I. Sensor Data Synchronization and TSDF / ESDF Voxel Map Construction The system synchronously receives depth images, odometry pose, and IMU inertial data and performs time alignment to obtain the camera pose in the world coordinate system. Compared with single-frame depth observation The depth observation ray is projected onto a three-dimensional discrete voxel mesh, where the side length of a single voxel is... Each voxel Includes occupied state TSDF values ESDF Euclidean distance Zone labels Four types of attributes. Perform TSDF voxel iterative updates, using the TSDF update calculation formula: ; In the formula This represents the TSDF value of the (i)th voxel at time (t). This represents the TSDF value of the voxel at the previous time step. Indicates the historical cumulative weight. This indicates the update weight corresponding to the current depth observation. This represents the voxel truncation symbol distance obtained from the depth observation of the current frame. Indicates the TSDF cutoff distance. This indicates that the calculation results will be restricted to... Within the range, multi-frame depth observations are continuously fused to reduce single-frame depth noise interference. The ESDF range field is incrementally solved based on the updated TSDF mesh, using the ESDF distance calculation formula: ; In the formula Representative voxels Euclidean distance to the nearest obstacle To occupy the entire set of voxels, For the position of the obstacle object, This represents the three-dimensional Euclidean distance between a voxel and an obstacle voxel. It sets the safe passage distance threshold. ,like The voxel is determined to be impassable; the free space set is defined as follows: Unknown frontier voxel set , Representative voxels After completing the spatial state partitioning of the 26-neighborhood voxels in 3D, extract the set of voxels whose states within the map update bounding box have been flipped. This serves as the input data for subsequent incremental topology decomposition.

[0034] II. Incremental 3D Connected Component Topology Decomposition Based on Zone State Machine Perform hierarchical mesh 3D connected component labeling (CCL) on the free voxels within the updated bounding box, and extract the set of center points of all free space connected domains in the current period. Retrieve the set of historical connected component center points from the previous period. Set a fixed matching distance threshold traverse each central point in the current period Calculate the Euclidean distance with all historical center points. Based on the matching results, Zone states are divided into three categories: expanded stable, region merging, and newly formed, achieving topological memory and local updates without the need for global clearing and reconstruction of the topology graph. If a single center point in the current period matches only one historical center point, it is determined to be in an expanded stable state. It directly inherits the center anchor point, connected edge set, and unknown voxel statistics of the historical Zone, and only updates the voxel occupancy quantity and boundary information of the current region. If a single center point in the current period matches multiple historical center points, it is determined to be in a region merging state. The Zone node set is updated with the current CCL connected component results, and only local path recalculation and repair are performed on connected edges of Zones with associations. If a current center point does not match any historical center points, it is determined to be a newly formed Zone. A new topology node is assigned, and the search for connected edges of surrounding adjacent Zones is started. During the CCL connected component extraction stage, micro-connected component filtering is performed simultaneously. If the total number of sampled voxels in the connected component is lower than the minimum voxel threshold, it is not registered as a valid Zone node. The unknown region filtering threshold is calculated by combining the voxel grid size, CCL sampling step size, and minimum unknown scale coefficient to eliminate scattered small regions with no exploration value and reduce redundant nodes in the topology graph. After processing all Zones, output the complete free space Zone set. Original connectivity relationships with Zone.

[0035] III. Extraction and Accessibility Verification of Adjacent ZonePortals For any two connected zones, use the local A* path search algorithm to find a collision-free path between the two zones. Take the midpoint of the path coordinate sequence As a representative point of the Portal, this Portal is stored in the built-in Portal set `portals_` of the two Zone nodes, completing the Zone node structure expansion and upgrading from a single center point representation to a composite spatial representation structure with a center point and multiple Portal channels. Portal accessibility verification is performed using the ESDF distance field. The ESDF distance value of the corresponding voxel of the Portal is read; if the distance is lower than the safe access threshold... The Portal is directly removed. Zones without valid Portal records default to using their own regional center point as the fallback candidate node for global planning. The Portal fully retains real passable geometric information such as doorways and corridors, making up for the lack of path accessibility caused by the abstract space of a single center point.

[0036] IV. Global Topology Planning of Mixed Granularity GTSP / ATSP under Local Viewpoint Constraints Using the real-time coordinates of the drone Set a fixed local view radius as a reference. Iterate through all zones, dividing them into near zones and far zones, satisfying... The Zone is divided into near-zones, and all internal Portal nodes are fully utilized to participate in the Generalized Traveling Salesman Service (GTSP) modeling; satisfying The zone is divided into far zones, and only the zone center point is used as the low-dimensional macroscopic guiding node. Zones without effective portals in the near zone also use their center points for solution. A threshold is set to increase the number of near zone portals. Count the total number of Portals in the near-field region that participated in the solution. ,when A Portal-GTSP model is constructed, and the Noon-Bean transformation is used to convert the GTSP model to an ATSP model. The LKH heuristic solver is then called to calculate the mixed Zone and Portal access sequence. Representative Portals are selected based on two criteria: the unknown volume of the Zone and the distance between the Portal and the UAV. If the total number of Portals still exceeds the threshold after selection, all Portal nodes are discarded, and the solution is reverted to the ATSP model that relies solely on the Zone center point. The connection cost between nodes is approximated using the Euclidean distance between Portals or center points and the path search cost, outputting a macroscopic access sequence that balances the accuracy of nearby channels with the efficiency of long-distance solutions.

[0037] V. Information-driven Local Active Perception SOP Viewpoint Ranking Calculation Extracting the set of all frontier clusters within the map under global zone access sequence constraints. The total number of voxels contained in a single cluster Defined as original information gain The information weights of frontiers of different scales are unified using a local normalization formula: ; In the formula, Representing the Normalized information gain of frontier clusters, Indicates the first A frontier cluster, This indicates the number of voxels contained in the frontier cluster. Indicates the first Information gain of a frontier cluster This represents the maximum information gain among the candidate frontier clusters currently participating in SOP sorting. To avoid extremely small constants with a denominator of zero, the normalization operation is performed only within the set of leading edges participating in the current sorting, ensuring that the system continuously focuses on locally relatively high-information detection regions. This is based on the maximum number of leading edge voxels in the current period. Calculate the dynamic greed coefficient The calculation is performed using the saturation function form: ; In the formula This represents the information greed coefficient for the current planning period. This represents the maximum greed coefficient. This represents the maximum number of front voxels in the current set of candidate fronts participating in SOP sorting. This represents the saturation adjustment constant. When the current local candidate front is large... As the size of the local front increases, the system tends to prioritize accessing regions with high information gain; when the size of the local front decreases... Automatic descent and gradual reverting to distance-first planning. The original physical transfer cost between viewpoints is constructed by combining flight path length, flight duration, and changes in fuselage yaw angle. The SOP ranking cost matrix is ​​reconstructed by introducing information gain discount, and the cost calculation formula is as follows: ; In the formula Indicates viewpoint Flying to the frontiers or candidate viewpoints The original physical cost can be composed of factors such as flight distance, flight time, and yaw variation; This represents the reconstruction cost after information gain discounting; This represents the dynamic greedy coefficient for the current planning cycle; Indicates the frontier of the target The normalized information gain. The higher the information gain at the target frontier, the better. The smaller the value, the more likely it is to be prioritized in the SOP sorting. Multiple layers of physical constraints are applied to all candidate viewpoints, with the constraint determination conditions as follows: In the formula, Indicates the first One candidate viewpoint, Indicates the ESDF safe distance at the candidate viewpoint. Indicates the safe travel distance for drones. Indicates the yaw angle corresponding to the candidate viewpoint. Indicates the reference yaw angle of the previous viewpoint or the current drone. Indicates the maximum permissible change in yaw. This means normalizing the angle difference to... Within the range, Indicates the target frontier cluster, Indicate candidate viewpoints At yaw angle The number of observable target frontal voxels This represents the minimum visible voxel count threshold. Candidate viewpoints that do not satisfy any constraint are not included in the SOP sorting, or their corresponding transition costs are set to infinity. After the effective viewpoints are filtered, the short-time domain local viewpoint access order is solved.

[0038] VI. Generation, Execution, and Iterative Processing of Dynamically Constrained Cubic B-Spline Trajectories Based on the local viewpoint sequence, UAV velocity and acceleration boundaries, and safe distance constraints, a cubic B-spline position trajectory is constructed. The fuselage yaw angle is optimized separately to generate a smooth yaw curve. During trajectory optimization, four-dimensional dynamic constraints—maximum linear velocity, maximum linear acceleration, maximum yaw angular velocity, and maximum yaw angular acceleration—are applied simultaneously. A fixed control point spacing is uniformly set to complete trajectory parameterization. After trajectory optimization, continuous control commands for position, velocity, acceleration, and yaw are issued to the flight controller at fixed intervals, executing short-time domain flight trajectories in real time. During trajectory execution, the map update status is continuously monitored. If a significant change occurs in the topology or a substantial change in the total global leading edge cluster, the current trajectory is immediately terminated, and the entire process of sensor mapping, incremental topology decomposition, portal extraction, hybrid-granularity global planning, and local viewpoint sorting is re-executed.

[0039] After a single complete iteration, all steps are executed in a loop, continuously updating the voxel map and generating new flight trajectories until the total number of unknown frontiers on the map is lower than a preset threshold or the environmental space coverage volume reaches a preset target standard, thus completing the UAV autonomous exploration process.

[0040] This invention is an autonomous exploration method for unmanned aerial vehicles (UAVs) based on local horizon adaptive perception and topology optimization. The overall system is composed of a sensor input unit, a voxel mapping unit, a state machine incremental spatial decomposition unit, a portal-based global guidance unit, an information-driven local active perception unit, and a trajectory generation and control unit. It fully covers the entire link of front-end perception mapping, incremental topology maintenance, hybrid granularity global path planning, information-driven local viewpoint sorting, and dynamic smooth trajectory generation.

[0041] I. Hardware and software operating environment This method can run on either a simulation platform or a physical drone, with unified hardware and parameter configurations in both environments. The simulation environment relies on the Ubuntu operating system, equipped with the ROS robot framework, RViz visualization tools, and a 3D voxel map module to complete the entire computation process. The physical device uses a quadcopter drone as the flight platform, equipped with an onboard depth camera or RGB-D sensor, and the onboard computing unit uses a multi-core CPU or a lightweight embedded GPU. The entire system relies only on two necessary input signals: depth images and onboard odometry. It does not access GPS positioning data throughout the process, making it suitable for operation in closed, unknown environments without satellite signals.

[0042] Standardize and fix hardware and map parameters: Set the maximum linear velocity of the drone to 2.0 m / s and the maximum linear acceleration to 3.0 m / s². 2 Maximum yaw rate: 1.5708 rad / s²; Maximum yaw acceleration: 1.5708 rad / s² 2 The airborne virtual depth camera has a resolution of 640×480, an effective depth detection range of 0.1m to 5.0m, a horizontal field of view of 1.5708rad, a vertical field of view of 1.2870rad, and camera intrinsic parameters. =320.0、 =320.0、 =320.0、 =240.0. The 3D voxel map uses a dual-layer distance field representation of TSDF and ESDF, with a fine voxel resolution of 0.1m and a coarse voxel resolution of 0.2m. The maximum distance measurement for ray projection is 5.0m, and the TSDF truncation distance is set to twice the resolution of a single voxel. The maximum cell size of the hierarchical grid is 5.0m, the sampling step size of the 3D connected component marker CCL is 2, and the minimum voxel scale factor for the unknown region is 0.1. The zone new and old center point matching distance threshold is 1.5m, the global safe passage distance is 0.7m, and the upper limit of the number of near-zone Portals participating in the solution is set. =80, the maximum value of the dynamic greedy coefficient. =0.85, saturation adjustment constant C=150.0, trajectory cubic B-spline control point spacing 0.5m.

[0043] II. Front-end voxel map construction and spatial state division The system simultaneously acquires depth images, odometry pose, and IMU inertial data, performs time alignment, and outputs the camera pose in the world coordinate system at the current moment. Compared with single-frame depth observation The depth observation ray is projected onto a discrete three-dimensional voxel grid, with each voxel... Record five types of attributes: 3D coordinates, occupancy status, TSDF value, ESDF Euclidean distance, and Zone region label. Perform multi-frame weighted TSDF fusion update using an iterative calculation formula: ; In the formula This represents the TSDF value of the (i)th voxel at time (t). This represents the TSDF value of the voxel at the previous time step. Indicates the historical cumulative weight. This indicates the update weight corresponding to the current depth observation. This represents the voxel truncation symbol distance obtained from the depth observation of the current frame. Indicates the TSDF cutoff distance. This indicates that the calculation results will be restricted to... Within the range, multi-frame observations are continuously fused to suppress single-frame depth imaging noise. The ESDF obstacle range field is incrementally solved based on the updated TSDF mesh, using the following formula: ; In the formula Represents the Euclidean distance from the voxel to the nearest obstacle voxel. , Represents the Euclidean norm of two points; if A voxel is deemed impassable if the safe passage distance is less than 0.7m. Three basic spatial sets are defined: free space, etc. Unknown frontier voxels , Represents a 26-neighborhood unit of a voxel in 3D; after each map update, the voxel set within the bounding box is extracted and flipped. , as input data for incremental topology decomposition.

[0044] III. Incremental Topology Decomposition Process Based on Zone State Machine For each free voxel within the updated bounding box, a hierarchical mesh 3D connected component labeling (CCL) is performed, and the set of all connected domain center points for the current period is extracted. Retrieve the set of historical center points from the previous period The algorithm calculates the Euclidean distance between each current center point and a historical center point, using a 1.5m matching threshold to classify topological changes into three categories: expansion stability, region merging, and newly formed Zones. Historical topological memory is preserved throughout the process, eliminating the need for global clearing and reconstruction of the connected graph. When the current center point matches only a single historical Zone, it is determined to be in an expansion stability state, directly inheriting historical region connectivity edges, Portal sets, and unknown voxel statistics, only updating the current region boundaries and voxel occupancy count. When the current center point matches multiple historical Zones, it is determined to be in a region merging state, only locally recalculating and repairing the connectivity edges of related regions. When the current center point has no matching historical nodes, it is determined to be in a newly formed Zone, allocating new topological nodes and searching for connectivity relationships with neighboring Zones. CCL extraction simultaneously performs micro-connectivity filtering; connected regions with fewer than 10 sampled voxels are directly discarded. Combining grid size, CCL sampling step size, and a minimum unknown scale coefficient of 0.1, it filters out scattered unknown regions with no detection value, reducing redundant nodes in the topological graph, and finally outputting a complete set of free space Zones and their connectivity relationships.

[0045] IV. Portal Channel Representative Point Extraction and Passage Verification When establishing a connection between any two Zones, a local A* path search is used to solve for a collision-free path, and the midpoint of the path coordinate sequence is taken as the representative point of the Portal. The data is stored in the built-in portals sets of both Zones, upgrading the Zone representation from a single center point to a structure with a center point and multiple passable Portal nodes. Portal security is verified using the ESDF distance field; if the Euclidean distance of the corresponding voxel is below the safety threshold of 0.7m, it is directly removed. Zones without valid Portals uniformly use the region's center point as a fallback candidate node for global planning. Portal nodes fully recreate the real-world geometry of doorways and narrow corridors, compensating for the path inaccessibility and detour defects caused by the abstract space of a single center point.

[0046] V. Local Horizon Mixed Granularity GTSP / ATSP Global Path Planning Using the real-time coordinates of the drone A local view radius is defined as a baseline. Based on the distance between the Zone and the UAV, two spatial regions are divided: near zone and far zone. Near zone Zones within the view radius fully utilize all internal Portals for Generalized Traveling Salesman Service (GTSP) modeling. Far zone Zones exceeding the view radius only retain the center point as a macroscopic guiding node. Near zone areas without effective Portals also use the center point for calculation. An upper limit of 80 Portal solutions is set, and the total number of participating Portals in the near zone is counted. , The Portal-GTSP model is directly constructed, and after being transformed into the ATSP model through the Noon-Bean transformation, the LKH heuristic solver is called to output a mixed access sequence of Zone and Portal. Representative Portals are selected based on two indicators: unknown volume of the region, Portal distance, and UAV distance. If the total number still exceeds the limit after selection, the solution is switched to ATSP solution only for the center point. The passage cost between nodes is calculated using Euclidean distance or A* path search cost, taking into account both the physical accuracy of the short path and the computational efficiency of the long-distance planning.

[0047] VI. Information-driven local viewpoint ranking calculation Extracting the entire front cluster set based on the global zone access sequence The total number of voxels contained in a single cluster is used as the original information gain, and a normalization formula is used to unify the frontier information weights of different scales: ; In the formula, Representing the Normalized information gain of frontier clusters, Indicates the first A frontier cluster, This indicates the number of voxels contained in the frontier cluster. Indicates the first Information gain of a frontier cluster This represents the maximum information gain among the candidate frontier clusters currently participating in SOP sorting. To avoid extremely small constants with a denominator of zero, normalization is performed only within the current candidate front, ensuring the system continuously identifies locally high-information detection regions. This is based on the maximum number of front voxels in the current period. Calculate the dynamic greed coefficient: ; In the formula This represents the information greed coefficient for the current planning period. This represents the maximum greed coefficient. This represents the maximum number of front voxels in the current set of candidate fronts participating in SOP sorting. This represents the saturation adjustment constant. When the current local candidate front is large... As the size of the local front increases, the system tends to prioritize accessing regions with high information gain; when the size of the local front decreases... Automatic descent, with planning gradually reverting to distance-first considerations. The original physical cost between viewpoints is constructed by combining flight path and fuselage yaw changes. The SOP sorting cost matrix is ​​reconstructed by introducing information gain: ; In the formula Indicates viewpoint Flying to the frontiers or candidate viewpoints The original physical cost can be composed of factors such as flight distance, flight time, and yaw variation; This represents the reconstruction cost after information gain discounting; This represents the dynamic greedy coefficient for the current planning cycle; Indicates the frontier of the target The normalized information gain. A higher frontier information gain corresponds to a lower access cost, making it easier to prioritize observations. Candidate viewpoints are simultaneously subject to multi-layered physical constraints for determination. ; In the formula, Indicates the first One candidate viewpoint, Indicates the ESDF safe distance at the candidate viewpoint. Indicates the safe travel distance for drones. Indicates the yaw angle corresponding to the candidate viewpoint. Indicates the reference yaw angle of the previous viewpoint or the current drone. Indicates the maximum permissible change in yaw. This means normalizing the angle difference to... Within the range, Indicates the target frontier cluster, Indicate candidate viewpoints At yaw angle The number of observable target frontal voxels This represents the minimum number of visible voxels threshold. Candidate viewpoints that do not satisfy any constraint are not included in the SOP sorting, or their corresponding transition costs are set to infinity. This filters out invalid viewpoints with collision risks, unreasonable orientations, or inability to observe unknown spaces, thereby reducing the computational cost of local planning.

[0048] VII. Generation and Iterative Process of Cubic B-Spline Smoothing Trajectory A cubic B-spline trajectory is generated based on the locally optimal viewpoint sequence. The position trajectory and fuselage yaw curve are optimized separately and independently. The spacing between trajectory control points is fixed at 0.5m. A maximum linear velocity of 2.0m / s and a maximum linear velocity of 3.0m / s are applied simultaneously throughout the optimization process. 2 Maximum linear acceleration, 1.5708 rad / s²; maximum yaw rate, 1.5708 rad / s². 2Maximum yaw angle acceleration is a four-dimensional dynamic constraint. After trajectory optimization, continuous control commands for position, velocity, acceleration, and yaw are sent to the flight controller at fixed intervals to execute short-time domain flight missions. During flight, the map topology and the total amount of global front changes are continuously monitored. If a significant change occurs, the current trajectory is immediately terminated, and the entire process of mapping, incremental topology decomposition, portal extraction, global planning, and local viewpoint sorting is re-executed. At the end of each iteration, all steps are executed cyclically until the total number of unknown fronts on the map is lower than a preset threshold or the 3D environment coverage volume reaches the set target, thus completing the autonomous exploration mission.

[0049] VIII. Standardized Implementation Testing Scenarios and Complete Implementation Steps 4.1 Test Scenario Configuration Three typical complex maze scenarios were selected for comparative verification. The scenario sizes were Maze1 (20×20×2), Maze2 (30×15×2), and Maze3 (20×20×4). Maze3 is a complex structure with two layers and multiple branches. The comparison algorithm used three existing mainstream UAV autonomous exploration methods: FUEL, FAEP, and FALCON. Each scenario was run independently 10 times, and the mean and standard deviation were calculated. The evaluation indicators included four items: total exploration time, cumulative flight distance, three-dimensional spatial coverage volume, and average flight speed.

[0050] 4.2 Complete Standardization Implementation Process 1. Initialize all configuration items for UAV flight parameters, voxel map parameters, field of view radius, Portal number threshold, and trajectory dynamics constraints, and clear historical Zone topology, front cluster, and trajectory cache data; 2. Start the depth camera and odometry data stream, collect sensor data in real time to complete coordinate alignment, iteratively update the TSDF and ESDF distance fields, and identify the state of the bounding box flipped voxels in the map. 3. Perform CCL three-dimensional connected component marking on the updated region, match the distance between the current connected domain center point and the historical center point, divide into three types of Zone states: expansion, merging, and new formation, locally repair the connected edges of the affected region, and filter out small and worthless connected domains; 4. For adjacent Zones with connectivity, call A* path search, extract the midpoint of the path as the Portal representative point and store it in the Portal set of the two ends, and remove unsafe Portals based on ESDF distance; 5. Divide the near and far zones based on the real-time location of the UAV, construct a hybrid granularity GTSP / ATSP optimization model, combine it with the Portal number threshold switching solution strategy, and output a global zone and Portal hybrid access sequence; 6. Extract all front clusters under global sequence constraints, calculate the number of front voxels, local normalized information gain, and dynamic greedy coefficient, reconstruct the viewpoint transition cost matrix, filter invalid viewpoints that do not conform to physical constraints, and solve the short-time domain optimal local viewpoint access order. 7. Generate a cubic B-spline smooth trajectory that satisfies all dynamic constraints based on the viewpoint sequence, and issue control commands to drive the UAV to fly; after the execution of a single short-time trajectory is completed, return to step 2 for iterative loop; 8. Continuously assess the remaining frontier quantity and spatial coverage volume. Once the termination condition is met, end the exploration process and output all experimental statistics.

[0051] IX. Analysis of Experimental Results Table 1: Performance Comparison of Multi-Algorithm Exploration in Three Maze Scenarios The comparison results of the three complex maze scenarios (Maze1, Maze2, and Maze3) in the table show that the method proposed in this invention achieves superior results in both exploration time and flight distance, the two core efficiency indicators, while maintaining a spatial coverage volume similar to other algorithms. Overall, the method of this invention can effectively shorten the time and flight path length required for UAVs to complete exploration of unknown spaces without significantly sacrificing coverage integrity. This demonstrates that the proposed incremental topology maintenance, local horizon hybrid granularity topology guidance, and information-driven local active perception mechanism can collectively improve exploration efficiency in complex environments.

[0052] In terms of average exploration time, the Proposed method achieved 75.53 s, 99.90 s, and 163.20 s in Maze1, Maze2, and Maze3, respectively, all lower than the three comparative methods FUEL, FAEP, and FALCON. Compared to FUEL, the proposed method reduced the average exploration time by approximately 42.05%, 36.10%, and 41.70% in the three scenarios; compared to FAEP, it reduced it by approximately 36.43%, 35.48%, and 40.68%; and compared to FALCON, it reduced it by approximately 4.40%, 7.64%, and 9.85%. Particularly in the complex two-layer maze scenario of Maze3, the proposed method further reduced the exploration time by 9.85% compared to FALCON, indicating that the method still has a good time efficiency advantage in scenarios with more complex topologies and richer spatial connectivity.

[0053] Regarding the average flight distance, the Proposed method achieved distances of 129.80 m, 163.01 m, and 275.86 m in the three scenarios, respectively, all of which were the best results among the compared algorithms. Compared to FUEL, the proposed method reduced the average flight distance in Maze1, Maze2, and Maze3 by approximately 24.06%, 19.06%, and 29.50%, respectively; compared to FAEP, it reduced the distance by approximately 5.43%, 12.43%, and 17.02%, respectively; and compared to FALCON, it reduced the distance by approximately 3.94%, 9.09%, and 12.49%, respectively. This demonstrates that by introducing a hybrid granularity of Portal channel representative points and local viewpoints at the global topology level, the proposed method can reduce detours and backtracking caused by the abstraction of a single region center point; simultaneously, information-driven local viewpoint ranking can reduce invalid access to low-value near-frontiers, thereby further compressing the cumulative flight distance.

[0054] In terms of coverage, the average coverage volume of the Proposed method in Maze1, Maze2, and Maze3 was 791.30 m³. 3 890.45 m 3 and 1595.15 m 3 The coverage volume of the proposed method is roughly on par with FUEL and FALCON, and is higher than that of FAEP in all three scenarios. Specifically, the proposed method improves the average coverage volume of FAEP by approximately 0.40%, 0.38%, and 2.05% in the three scenarios, respectively; while its coverage volume is slightly lower than that of FUEL and FALCON, the differences are small, with the largest difference not exceeding approximately 0.46%. Therefore, the experimental results show that the time and path advantages of the proposed method are not obtained by reducing the exploration area or reducing coverage integrity, but mainly by more stable topology maintenance, a more reasonable global access order, and more effective information-oriented local planning.

[0055] In summary, the method of this invention achieves the combined effects of "shorter exploration time, shorter flight distance, and sufficient coverage volume" in all three complex maze scenarios. Particularly in the Maze3 two-layer complex space, the method of this invention demonstrates significant time and path advantages compared to FUEL, FAEP, and FALCON, while maintaining near-maximum coverage volume. This verifies the effectiveness and practical value of state machine incremental topology maintenance, the hybrid granularity topology guidance mechanism under local horizon constraints, and the information-driven local active perception planning mechanism in autonomous exploration missions of UAVs in complex and unknown environments.

[0056] Reference Figure 1 This figure macroscopically illustrates the complete operational cycle and core module architecture of an UAV adaptive autonomous exploration method based on local horizon topology optimization. The process begins with the synchronous input of multi-source sensor data. The system integrates depth images and inertial odometry information, initializing and continuously updating a 3D voxel map containing truncated symbolic distance and Euclidean distance fields. Based on environmental perception, the system performs incremental spatial topology decomposition, dynamically maintaining the state machine of spatially connected regions, and extracting access portals between adjacent regions and performing security checks. Subsequently, the planning module performs hybrid-granularity global path planning based on local horizon constraints, determining the macroscopic region and portal access sequence. Under this global guidance, the system further combines frontier information gain and physical transfer cost to complete adaptive local viewpoint sequencing. Finally, the control module generates a smooth cubic B-spline flight trajectory under multidimensional dynamic constraints and issues it for execution. The system iterates through the above process until the coverage of unknown areas meets the preset exploration completion criteria.

[0057] Reference Figure 2 This diagram details the underlying logic of incremental construction and maintenance of the 3D spatial topology map and the extraction of channel features in connected regions. To address the massive computational overhead caused by frequent local map updates, the system adopts an incremental processing strategy, extracting only voxels whose states have flipped and performing 3D connected component labeling. After filtering out small invalid regions, it generates a set of center points for the current free space connected domain. By matching the distance between the current center points and historical periodic nodes, the system intelligently determines whether each region is in an expanding stable state, a region merging state, or a newly generated state. Based on this, it locally repairs or expands the topology network, retaining historical memory throughout to prevent path oscillations. On this topological framework, when adjacent nodes establish connectivity, the algorithm automatically calls local path search to extract the midpoint of the collision-free path as the Portal representative point, and combines distance field data to remove invalid Portal points that are close to the safety limit. This precise Portal extraction mechanism effectively compensates for the non-convex space pathfinding and detour defects caused by relying solely on region center points.

[0058] Reference Figure 3This paper focuses on analyzing the hierarchical planning process that balances computational efficiency and information gain. In the global planning phase, the system divides the local near-field and global far-field regions using the UAV's real-time location as the origin. For the near-field region within the field of view, all valid Portals are introduced as candidate nodes to ensure local reachability, while for the far-field region, only the center point is retained as a macroscopic guide. The system adaptively constructs a generalized or asymmetric traveling salesman model based on the node size, and outputs a hybrid access sequence through transformation and heuristic solving, successfully curbing the combinatorial explosion problem in complex environments. In the local planning phase, the algorithm extracts unknown frontier clusters under the guidance of the global sequence, extracts the total size of voxels in a single cluster, and performs normalization. Subsequently, it integrates dynamic greed coefficients, flight distance, and yaw changes to reconstruct a transfer cost matrix that balances physical energy consumption and information gain. After eliminating redundant viewpoints that do not meet camera field of view and safety constraints, it accurately outputs a high-yield, highly secure local viewpoint access sequence.

Claims

1. A UAV adaptive autonomous exploration method based on local horizon topology optimization, characterized in that, The complete execution steps are as follows: S1 performs time registration of depth images, odometry pose data and IMU inertial data, constructs TSDF voxel mesh and incrementally solves ESDF distance field, divides voxel occupancy states, and stores voxel connectivity labels, cutoff distances, cumulative weights and passage distance constraints. S2 filters voxels with state changes within the bounding box, extracts free space connected components through 3D connected component labeling and obtains the zone center point, performs threshold matching between the current center point and the historical center point, distinguishes the three types of zone states and performs corresponding topology update operations, and removes tiny connected components with too few voxels or too small volume. S3 constructs connecting edges between adjacent Zones, uses the local A* algorithm to solve for collision-free traversable paths, extracts the midpoint of the path as the Portal representative point, and records it in the Portal set of adjacent Zone nodes; combining path search collision constraints, free space filtering, and ESDF distance field to verify the Portal traversability, invalid Portal points are eliminated, and Zones without valid Portal records use their own center point as the planning node; S4 defines the local field of view based on the drone's position, dividing the Zone into a near zone and a far zone. The near zone uses Portal nodes, while the far zone only retains the center point. If the number of Portals is within the limit, a Portal-GTSP model is constructed and converted into an ATSP model to solve the path; if the limit is exceeded, the Portals are simplified and degenerated into center point planning, and a mixed access sequence is output. S5 extracts the environmental frontier clusters and calculates the information gain. After normalization, it constructs a cost matrix by combining flight distance and attitude cost. It prioritizes high information frontier viewpoints, eliminates candidate points that violate flight constraints, and solves the local viewpoint access order. S6 combines the viewpoint sequence with UAV dynamic constraints to construct a cubic B-spline smooth trajectory and optimize the yaw angle curve. It periodically issues flight control commands and iteratively executes S1 to S6 until the space exploration coverage is achieved.

2. The UAV adaptive autonomous exploration method based on local horizon topology optimization according to claim 1, characterized in that, In step S1, the TSDF voxel update uses a weighted iterative pruning calculation, and the calculation formula is as follows: ； In the formula This represents the TSDF value of the (i)th voxel at time (t). This represents the TSDF value of the voxel at the previous time step. Indicates the historical cumulative weight. This indicates the update weight corresponding to the current depth observation. This represents the voxel truncation symbol distance obtained from the depth observation of the current frame. Indicates the TSDF cutoff distance. This indicates that the calculation results will be restricted to... Within the range.

3. The UAV adaptive autonomous exploration method based on local horizon topology optimization according to claim 1, characterized in that, The core calculation formula for the frontier information gain normalization and reconstruction cost matrix in step S5 is as follows: In the formula, Representing the Normalized information gain of frontier clusters, Indicates the first A frontier cluster, This indicates the number of voxels contained in the frontier cluster. Indicates the first Information gain of a frontier cluster This represents the maximum information gain among the candidate frontier clusters currently participating in SOP sorting. To avoid extremely small constants with a denominator of zero, Indicates viewpoint Flying to the frontiers or candidate viewpoints The original physical cost, This represents the reconstruction cost after information gain discounting; This represents the dynamic greedy coefficient for the current planning cycle; Indicates the frontier of the target Normalized information gain.

4. The UAV adaptive autonomous exploration method based on local horizon topology optimization according to claim 1, characterized in that, In step S2, the matching of the old and new center points of the Zone uses the Euclidean distance rule and sets a fixed matching distance threshold. For any Zone center point in the current period, traverse all historical Zone center points from the previous period. The Euclidean distance between any two points is less than 1. If a zone is determined to be an evolution of the same connected region, it directly inherits the connected edges, Portal set, and unknown voxel statistics of the historical Zone. If a single Zone center point matches multiple historical Zone center points in the current period, it is determined to be a region merging state. Only the connected edges of the Zones with association are locally recalculated. If the current center point has no matching historical nodes, it is determined to be a new Zone. A new topology node is assigned and the search for surrounding connected edges is started.

5. The UAV adaptive autonomous exploration method based on local horizon topology optimization according to claim 1, characterized in that, Step S3, the Portal representative point extraction process, is completed synchronously with the Zone connected edges. When two Zones establish a connection, a local A* path search is started. After the path traversal is completed, the middle position of the path coordinate sequence is taken as the Portal representative point and stored in the built-in Portal set of the two Zone nodes respectively. In the subsequent global GTSP planning stage, the Portal set is directly read as the candidate node in the cluster. For Zones that have not generated a Portal, the regional center point is used as the fallback node. The Portal stores the actual geometric position of the passageway.

6. The UAV adaptive autonomous exploration method based on local horizon topology optimization according to claim 1, characterized in that, Step S4 divides the region into near and far zones based on the horizon radius. Near zones prioritize using selected Portal nodes for GTSP calculations, while far zones only use zone center points for ATSP calculations. A threshold is set for the number of near-zone Portals. If the number exceeds the threshold, Portals are selected based on the distance between the Portal and the UAV's current position, the number of unknown voxels in the zone, and the connectivity cost index. If the number still exceeds the threshold after selection, center point planning is switched to control the problem size. The GTSP model is converted to an ATSP model using the Noon-Bean transformation, and the LKH solver is called to obtain the traversal order of the zones and Portals. This macro-sequence only limits the traversal order; the flight trajectory is refined by the local viewpoint planning module, and hierarchical planning reduces the solution dimension.

7. The UAV adaptive autonomous exploration method based on local horizon topology optimization according to claim 1, characterized in that, Step S5: The dynamic greedy coefficient is adaptively adjusted according to the number of maximum voxels at the local front. The dynamic greedy coefficient is calculated using a saturation function. In the formula This represents the information greed coefficient for the current planning period. This represents the maximum greed coefficient. This represents the maximum number of front voxels in the current set of candidate fronts participating in SOP sorting. Represents the saturation adjustment constant; Step S5 sets up multi-layer physical constraints to filter candidate viewpoints. Candidate viewpoints must simultaneously satisfy the following conditions: ESDF distance is greater than the safe passage distance, the difference in yaw angle between adjacent viewpoints is less than the maximum yaw change, and the leading edge region is within the camera's field of view. If any constraint is not met, the corresponding transfer cost is set to infinity, and the viewpoint will no longer participate in the SOP sorting calculation, thus eliminating invalid viewpoints in advance.

8. The UAV adaptive autonomous exploration method based on local horizon topology optimization according to claim 1, characterized in that, Step S6 uses cubic B-splines to complete trajectory parameterization, uniformly sets fixed control point spacing, and simultaneously applies four-dimensional dynamic constraints of maximum linear velocity, maximum linear acceleration, maximum yaw rate, and maximum yaw rate during the trajectory optimization stage. The position trajectory and yaw trajectory are optimized separately to generate smooth curves. If the map update causes a significant change in the Zone topology or a significant change in the total number of leading clusters during a single trajectory execution, the current trajectory is immediately terminated and global path planning and local viewpoint sorting are re-executed to dynamically respond to spatial structure changes caused by environmental perception.

9. The UAV adaptive autonomous exploration method based on local horizon topology optimization according to claim 1, characterized in that, Step S2 adds a small connected component filtering mechanism after the three-dimensional connected components are marked. When the total number of free voxels in a connected component is less than the minimum voxel threshold, it is not registered as a valid Zone node. At the same time, the unknown connected component filtering threshold is calculated based on the voxel grid size, CCL sampling step size, and minimum unknown scale coefficient to remove scattered unknown regions.

10. The UAV adaptive autonomous exploration method based on local horizon topology optimization according to claim 1, characterized in that, This method does not rely on GPS signals throughout the process, but only on an airborne depth camera, odometry and IMU to achieve perception and positioning. The algorithm is deployed on a small quadcopter drone embedded platform. The voxel map adopts a dual-layer structure with a fine resolution of 0.1m and a coarse resolution of 0.2m. The observation distance, TSDF cutoff distance, grid size and sampling step size are all adaptively adjusted according to the scene and computing power.

Images