An unmanned aerial vehicle layered exploration method based on gravity mechanism view point generation

Abstract

The application discloses a kind of unmanned plane layered exploration methods based on gravity mechanism view point generation;The method uses layered mode, and different gravity mechanisms are proposed for fine scene and broad scene:When broad scene, unmanned plane uses interactive lateral gravity mechanism to update view point, so that view point produces distance unmanned plane itself body closer, reduce unnecessary long distance maneuver because view point distance unmanned plane is far;When fine scene, narrow lane gravity mechanism is used to sample and generate view point, so that view point produces smaller radius, and then make unmanned plane have longer track planning, reduce inefficient maneuver of advance-back, finally make unmanned plane complete exploration task in a kind of high precision, high efficiency.

Description

A Layered Exploration Method for UAVs Based on Gravity Mechanism Viewpoint Generation Technical Field

[0001] This invention belongs to the field of autonomous unmanned exploration technology for unmanned aerial vehicles (UAVs) in unknown environments, and particularly relates to a UAV layered exploration method based on gravity mechanism viewpoint generation. Background Technology

[0002] Autonomous exploration by unmanned aerial vehicles (UAVs) has shown great potential in emergency rescue and air-ground coordination, and its progress has attracted the attention of many researchers in recent years. Generally speaking, autonomous exploration by UAVs refers to the process by which UAVs, within a designated area, autonomously explore and scan the environment by sequentially performing processes such as acquiring environmental perception data, incremental mapping and environmental representation, trajectory problem modeling, trajectory generation, and executing trajectory actions. This allows them to reconstruct the scene of the designated area through mapping, which can then be used for other tasks such as path planning by UAVs and unmanned vehicles.

[0003] However, existing methods still have limitations in multi-scale environments, failing to adapt well to changes in environmental size, thus significantly reducing exploration efficiency. Specifically, existing methods generally limit the autonomous exploration environment of UAVs to relatively wide areas, where the efficiency and safety of UAVs performing exploration tasks are well guaranteed. However, in multi-scale scenarios that combine wide and narrow scenes, existing methods face a conflict between efficiency and precision: when the obstacle expansion coefficient is high, the UAV exploration efficiency is high and the planned trajectory is smooth, but it cannot enter narrow scenes; when the expansion coefficient is low, the UAV exploration accuracy is guaranteed and it can successfully enter narrow areas, but it exhibits more hesitant and repetitive maneuvers. Summary of the Invention

[0004] In view of this, the present invention provides a UAV layered exploration method based on gravity mechanism viewpoint generation, enabling UAVs to complete exploration tasks in an innovative gravity mechanism viewpoint generation manner in unknown multi-scale environments with high precision and efficiency. The method acquires sensor pose and depth images from the UAV's forward-looking depth camera at the current moment, incrementally constructing a global map; it updates all frontiers within the global map using an overlapping region judgment mechanism; it determines whether the scene scale is fine-grained or wide-grained; based on whether the scene scale is fine-grained or wide-grained, it uses a narrow-path gravity mechanism to sample and generate viewpoints / updates viewpoints using an interactive lateral gravity mechanism; it generates an ATSP cost matrix using the current architecture viewpoint, solves the matrix to obtain the UAV waypoint sequence; and it calculates the number of remaining frontiers, continuing the loop if the number is not zero, and ending the exploration if the number is zero.

[0005] A layered exploration method for UAVs based on gravity mechanism viewpoint generation includes:

[0006] Step 1: The UAV acquires the sensor pose T and depth image I using the forward-looking depth camera. depth Incremental construction of global map M globe ;

[0007] Step 2: Using the overlapping region detection mechanism, update all frontiers in the global map at the current time t, and then obtain the frontier set F. t ;

[0008] Step 3: Calculate the frontier set F t The loop iterates through the remaining number of leading edges. If the return value is not zero, the loop continues to the next step. If the return value is zero, the loop is considered to have reached the end point and ends.

[0009] Step 4: Determine the scene scale to confirm whether the current scene is a detailed scene or a wide scene;

[0010] Step 5: For detailed scenes, use the narrow gravity mechanism to sample and generate viewpoints; for wide scenes, use the interactive lateral gravity mechanism to update viewpoints.

[0011] Step 6: Generate the ATSP cost matrix using the viewpoint generated in Step 5. cost Solve the matrix to obtain the UAV waypoint sequence P. waypoints Return to step 1.

[0012] Preferably, in step 1, after acquiring the sensor pose T and depth image I... depth Then, for the depth image Id epth Each of them has pixel coordinates (u n v n ), depth value Z n Perform the following operations on the pixels:

[0013]

[0014] Among them, P w for That is, the coordinates of the pixel in the world coordinate system, T is the current pose of the sensor obtained, and K is the intrinsic parameter matrix of the camera;

[0015] Repeat the above steps for each depth image, converting each pixel into a point in the world coordinate system, and accumulate these point clouds to form a complete 3D point cloud M at the current time t. t ;

[0016] Next, using point cloud registration methods, a global map M is incrementally constructed by utilizing the complete 3D point cloud at the current moment and the 3D point clouds at all past moments. globe .

[0017] Preferably, in step 2, the method for updating all frontiers in the global map at the current time t includes:

[0018] First, obtain the complete 3D point cloud M at the current time t. t Convert to voxel map V t Then, based on the occupancy status, the voxel states are divided into occupied, unknown, or free; using a growth algorithm, voxels adjacent to unknown voxels in the same region are clustered into front f groups. k ; where f k It contains all the constituent voxels (cells), the mean position of the cluster, and the coordinates (x, y, z) of the cluster bounding box.

[0019] Preferably, in step 2, the frontier set F is obtained. t The methods include:

[0020] Voxel map V detected at the current moment t Boundary and the previous time set F t-1 For all leading bounding box coordinates (x, y, z), determine the overlapping region. If (x, y, z) is contained within the voxel map V... t Within the boundary, delete this frontier; and so on, calculate the voxel map V at the current time. t The frontier f k Add it to the previous time-series set F that was just processed. t-1 In this process, we obtain the frontier set F at the current moment. t .

[0021] Preferably, in step 4, the method for determining the scene scale includes:

[0022] The set AF comprises all frontiers deleted at the current time t. t Calculate the mean position of the cluster and the coordinates (x, y, z) of the cluster bounding box; use the three most recently dormant abandoned bounding boxes AF. t AF t-1 and AF t-2 As a judgment sequence, it determines whether the current scene is a fine scene or a broad scene:

[0023] The first step is to determine the three sets of abandoned boundaries AF. t AF t-1 and AF t-2 Check if all dimensions meet the narrowness requirement. If they do, proceed to step two to determine if it is a narrow scene. If any set does not meet the requirement, determine it is a wide scene and proceed to step 5.

[0024] The second step is to determine the abandoned boundary AF in three consecutive updates.t AF t-1 and AF t-2 Arrangement characteristics:

[0025]

[0026] in:

[0027] m t =AF t .mean_position

[0028] Judgment mechanism:

[0029]

[0030] This is used to determine the scale of the scene.

[0031] Preferably, in step 4, whether the narrow requirement is met is: abandoned boundary AF t AF t-1 and AF t-2 Three-dimensional coordinates of arbitrary voxels and AF t Is the average distance between cluster positions less than a set threshold?

[0032] Preferably, in step 5, based on the distinction between detailed and wide scenes, the viewpoint generation method is determined, including:

[0033] For detailed scenes, a narrow-channel gravity mechanism is used to sample and generate the viewpoint F. t .f k .viewpoints, these viewpoints are composed of n sampling points p i Composition; sampling point p i Expressed in polar coordinates, with the origin at the average position of the front edge F. t .f k .mean_position:

[0034]

[0035] Where the polar diameter r i The distribution conforms to:

[0036]

[0037] polar angle The distribution conforms to:

[0038]

[0039] If sampling point p i If the voxel overlaps with the occupied state voxel or the unknown state voxel, then resample.

[0040] In open scenes, the viewpoint F is updated using an interactive lateral gravity mechanism. t .f k .viewpoints, these viewpoints are composed of n sampling points p i Composition; sampling point p i Expressed in polar coordinates, with the origin at the average position of the front edge F. t .f k .mean_position:

[0041]

[0042] Where the polar diameter r i The distribution conforms to:

[0043]

[0044] polar angle The distribution conforms to:

[0045]

[0046] Where θ is the current average position of the frontier F t .f k .mean_position points to the drone's own coordinates p ego and the current frontier average position F t .f k .mean_position points to the sampling point p i The angle between them; all of the above r are fixed values.

[0047] Preferably, in step 6, the ATSP cost matrix is ​​generated. cost The methods include:

[0048] Using the current set F t Generate the ATSP cost matrix. cost , where the element Matrix cost (k1, k2) represents the cost from the k1th front to the k2th front. Matrix cost (0, k1) represents the cost of the drone's position reaching the k1-th front edge, Matrix cost (k1, 0) represents the cost from the k1th front edge to the drone's location, and it is generated as follows:

[0049] The elements in the first column are:

[0050] Matrix cost (k, 0) = 0, 0 ≤ k ≤ n

[0051] The elements in the first row are:

[0052]

[0053] Where n represents the number of viewpoints, p ego p represents the drone's own coordinates. k For F t The average position F of the k-th frontal viewpoint f t .f k .viewpoints_meanposition, Let p be the velocity vector of the UAV, length(p) ego p k ) represents the path length from the UAV to the k-th leading edge;

[0054] In addition, the ATSP cost matrix cost The elements of the n×n square matrix excluding the first row and first column are:

[0055]

[0056] Preferably, in step 6, the LKH solver is used to solve the matrix to obtain the leading edge f visit sequence, which is then used as the UAV waypoint sequence P. waypoints And execute.

[0057] The present invention has the following beneficial effects:

[0058] This invention addresses the efficiency / accuracy conflict faced by unmanned aerial vehicles (UAVs) in autonomous exploration across multi-scale scenarios by designing a layered exploration method based on gravity-based viewpoint generation. This method employs a layered approach, proposing different gravity mechanisms for fine and wide scenarios: In wide scenarios, the UAV uses an interactive lateral gravity mechanism to update the viewpoint, making the viewpoint closer to the UAV's own body and reducing unnecessary long-distance maneuvers due to the viewpoint being too far away; In fine scenarios, a narrow-channel gravity mechanism is used to sample and generate the viewpoint, resulting in a smaller viewpoint radius and allowing the UAV to plan longer flight paths, reducing inefficient forward-backward maneuvers. Ultimately, this enables the UAV to complete exploration tasks with high precision and efficiency. Attached Figure Description

[0059] Figure 1 is a flowchart of the UAV layered exploration method based on gravity mechanism viewpoint generation according to the present invention.

[0060] Figure 2 is a schematic diagram of the narrow-channel gravity mechanism sampling and viewpoint generation algorithm of the present invention;

[0061] Figure 3 is a schematic diagram of the interactive lateral gravity mechanism update viewpoint algorithm of the present invention. Detailed Implementation

[0062] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0063] Step 1: The UAV acquires the sensor pose T and depth image I using the forward-looking depth camera. depth Incremental construction of global map M globe ;

[0064] Step 2: Use the overlapping region detection mechanism to update all leading edges F in the global map at the current time t. t ;

[0065] Step 3: Calculate the number of remaining frontiers F t If the return value of .size() is not zero, the next step is executed. If the return value is zero, the end point is reached and the loop ends.

[0066] Step 4: Based on the discarded AF data from the three most recent moments t AF t-1 and AF t-2 Determine the scene scale to confirm whether the current scene is a detailed scene or a wide scene;

[0067] Step 5: For detailed scenes, use a narrow-path gravity mechanism to sample and generate the viewpoint F. t .f k .viewpoints, in the case of a wide scene, use an interactive lateral gravity mechanism to update the viewpoint F. t .f k .viewpoints;

[0068] Step 6: Generate the ATSP cost matrix using the current architecture viewpoint. cost Solve the matrix to obtain the UAV waypoint sequence P. waypoints Next, continue with step one;

[0069] Furthermore, in step 1, after acquiring the sensor pose T and depth image I... depth Then, for depth image I depth Each of them has pixel coordinates (u n v n ), depth value z n Perform the following operations on the pixels:

[0070]

[0071] Among them, P w for That is, the coordinates of the pixel in the world coordinate system, T is the current pose of the sensor, and K is the intrinsic parameter matrix of the camera.

[0072] Repeat the above steps for each depth image, converting each pixel into a point in the world coordinate system, and accumulate these point clouds to form a complete 3D point cloud M at the current time t. t .

[0073] Next, using point cloud registration methods, a global map M is incrementally constructed by utilizing the complete 3D point cloud at the current moment and the 3D point clouds at all past moments. globe .

[0074] In step 2, the complete 3D point cloud M at the current time t is first... t Convert to voxel map V t Then, based on the occupancy state, the voxel states are divided into occupied (with obstacles), unknown, and free (known to be free of obstacles), and all frontier f values ​​detected at the current moment are included. k Stored in F t In the middle, F t It is referred to as the set of all frontiers within the global map at the current time t.

[0075] Frontier f k Detection method: Using a growth algorithm, voxels adjacent to unknown voxels in the same region are clustered into a leading edge f. k f k It contains all the constituent voxels (cells), the mean position of the cluster, and the coordinates (x, y, z) of the cluster bounding box.

[0076] Frontier set F t Detection method: The voxel map V detected at the current moment t Boundary and the previous time F t-1 All F in t-1 .f k Use the bounding box coordinates (x, y, z) to determine overlapping regions. If (x, y, z) is contained within the voxel map V... t Delete this F within the boundary. t-1 .f k Based on cutting-edge detection methods, calculate the voxel map V at the current time. t The frontier f k Add it to the F that was just processed t-1 In this process, we obtain the frontier set F at the current moment. t The above F t-1 .fk The definition is the frontier set F at time t-1. t-1 The k-th frontier f k .

[0077] All F deleted at current time t t-1 .f k .cells form a collection AF t Calculate the mean position of the cluster and the coordinates (x, y, z) of the cluster bounding box.

[0078] In step 4, the three sets of abandoned boundary AFs that recently entered a dormant state are used. t AF t- 1 and AF t-2 As a judgment sequence, it confirms whether the current scene is a detailed scene or a broad scene.

[0079] The core idea of ​​the layered exploration method is to divide the exploration scene into a fine-scale scene and a wide-scale scene. After initialization, the drone exploration always runs in a wide-scale scene exploration mode, during which an interactive lateral gravity mechanism is used to generate the boundary viewpoint. When the boundary between the fine-scale scene and the wide-scale scene is detected—that is, when we propose a switching boundary—it will switch to a narrow-scale area exploration mode.

[0080] The transition boundary at the junction of detailed and wide scenes has unique morphological and sequence features. The algorithm uses three sets of abandoned boundaries (AFs) that have recently entered a dormant state. t AF t-1 and AF t-2 Make a judgment:

[0081] The first step is to determine the three sets of abandoned boundaries AF. t AF t-1 and AF t-2 Do all dimensions meet the narrow requirements, according to AF? t For example:

[0082] ||AF t .cell-AF t .mean_position|| <r narrow

[0083] Where AF t .cell is AF t 3D coordinates of any voxel, AF t .mean_position is AF t The average position of the cluster.

[0084] Three sets of abandoned boundaries AF t AF t-1 and AF t-2If the dimensions all meet the narrowness requirement, proceed to the second step to determine if it is a narrow scene; otherwise, it is determined to be a wide scene, and the second step is not performed.

[0085] The second step is to determine the abandoned boundary AF in three consecutive updates. t AF t-1 and AF t-2 Arrangement characteristics:

[0086]

[0087] in:

[0088] m t =AF t .mean_position

[0089] Judgment mechanism:

[0090]

[0091] This allows us to determine the scale of the scene.

[0092] Furthermore, in step 5, the viewpoint generation method is determined based on the judgment of detailed scene / wide scene.

[0093] The viewpoint is the frontier F t .f k The way nearby locations accessible to drones are generated plays a crucial role in the efficiency of drone exploration.

[0094] For detailed scenes, as shown in Figure 2, a narrow-path gravity mechanism is used to sample and generate the viewpoint F. t .f k .viewpoints, these viewpoints are composed of n sampling points p i Composition. Sampling point p i Expressed in polar coordinates, with the origin at the average position of the front edge F. t .f k .mean-position:

[0095]

[0096] Where the polar diameter r i The distribution conforms to:

[0097]

[0098] polar angle The distribution conforms to:

[0099]

[0100] If sampling point p i If the voxel overlaps with the occupied state voxel or the unknown state voxel, then resample.

[0101] In a wide scene, as shown in Figure 3, the viewpoint F is updated using an interactive lateral gravity mechanism. t .f k .viewpoints, these viewpoints are composed of n sampling points p i Composition. Sampling point p i Expressed in polar coordinates, with the origin at the average position of the front edge F. t .f k .mean-position:

[0102]

[0103] Where the polar diameter r i The distribution conforms to:

[0104]

[0105] polar angle The distribution conforms to:

[0106]

[0107] Where θ is the current average position of the frontier F t .f k .mean-position points to the drone's own coordinates p ego and the current frontier average position F t .f k .mean_position points to the sampling point p i The angle between them. All of the above r values ​​are fixed values.

[0108] After sampling, calculate F for all viewpoints along this front. t .f k The average value of the viewpoints is F. t .f x .viewpoints_meanposition.

[0109] Furthermore, in step 6, the F of the current architecture is utilized. t Generate the ATSP cost matrix (containing the average position of n front viewpoints). cost , where the element Matrixcost (k1, k2) represents the cost from the k1th front to the k2th front. Matrix cost (0, k1) represents the cost of the drone's position reaching the k1-th front edge, Matrix cost (k1, 0) represents the cost from the k1th front edge to the drone's location, and it is generated as follows:

[0110] The elements in the first column are:

[0111] Matrix cost (k,0)=0,0≤k≤n

[0112] The elements in the first row are:

[0113]

[0114] Where n represents the number of viewpoints, p ego p represents the drone's own coordinates. k For F t The average position F of the k-th frontal viewpoint f t .f k .viewpoints_meanposition, Let p be the velocity vector of the UAV, length(p) ego ,p k ) represents the path length from the UAV to the k-th leading edge;

[0115] In addition, the ATSP cost matrix cost The elements of the n×n square matrix excluding the first row and first column are:

[0116]

[0117] All the length() calculations above, i.e., the length between two points, are performed using the A* algorithm.

[0118] Next, the LKH solver is used to solve the matrix to obtain the leading edge f-visit sequence, which is then used as the UAV waypoint sequence P. waypoints And execute.

[0119] After execution is complete, continue executing sequentially from step one.

[0120] A multi-scale field of 50×50m was selected for testing the improved algorithm, and the maximum flight speed v was set. max=2 (m / s), maximum acceleration a = 2 (m / s), maximum angular velocity α = 0.9 (rad / s). In actual tests, when the drone flies in a narrow passage, it flies a longer distance towards the next boundary. The intuitive effect is that the boundary sequence in a narrow passage exerts a greater gravitational pull on the drone, greatly reducing the back-and-forth flight phenomenon in narrow passages; in a wide scene, each time the boundary is updated, the algorithm generates a viewpoint of the boundary within a certain radius centered on the drone in a lateral viewpoint optimization manner, making the generated viewpoint easy for the drone to scan with minimal maneuvering cost at its current position and speed.

[0121] The final test results show that this method has significant advantages over existing methods in key indicators.

[0122] Algorithm path total length (m) Total exploration time (s) Invention 794.36 22.1 Existing method 883.77 13.9 surface

[0123] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A layered exploration method for unmanned aerial vehicles (UAVs) based on viewpoint generation using gravity mechanisms, characterized in that, include: Step 1: The drone acquires the sensor pose using a forward-looking depth camera. and depth images Incremental construction of global map Step 2: Use the overlapping region detection mechanism to update the current time. All frontiers within the global map are then used to obtain the set of frontiers. Step 3: Calculate the frontier set The remaining number of frontiers is counted. If the return value is not zero, the next step is executed. If the return value is zero, it is considered that the endpoint has been reached and the loop ends. Step 4: Determine the scene scale to confirm whether the current scene is a fine scene or a wide scene. Step 5: If it is a fine scene, use the narrow-path gravity mechanism to sample and generate viewpoints. If it is a wide scene, use the interactive lateral gravity mechanism to update viewpoints. Step 6: Generate the ATSP cost matrix using the viewpoints generated in Step 5. Solve the matrix to obtain the UAV waypoint sequence. Return to step 1; in step 2, update the current time. The method for identifying all frontiers within the global map includes: first, setting the current time... Complete 3D point cloud Convert to voxel map Then, based on the occupancy status, the voxel states are divided into occupied, unknown, or free; using a growth algorithm, voxels adjacent to unknown voxels in the same region are clustered into frontiers. ;in, It contains all the constituent voxels Cluster average position and cluster bounding box coordinates In step 2, the frontier set is obtained. The method includes: using the voxel map detected at the current moment Boundary and previous time set Coordinates of all front bounding boxes Perform overlapping region judgment, if Included in voxel maps Within the boundary, delete this frontier; and so on, calculate the voxel map at the current time. China's frontier Add it to the set of the previous time step that was just processed. In this process, we obtain the frontier set at the current moment. In step 4, the method for determining the scene scale includes: the current moment. The set of all deleted frontier components Calculate its average cluster position and cluster bounding box coordinates Use the three sets of abandoned boundaries that recently entered a dormant state. and As a judgment sequence, determine whether the current scene is a detailed scene or a wide scene: The first step is to judge three sets of abandoned boundaries. and If all dimensions meet the narrowness requirement, proceed to step two to determine if it is a narrow scene; if any set does not meet the requirement, it is determined to be a wide scene, and step 5 is executed; step two: determine abandoned boundaries in three consecutive updates. and Arrangement characteristics: in: Judgment mechanism: This is used to determine the scale of the scene.

2. The UAV layered exploration method based on gravity mechanism viewpoint generation as described in claim 1, characterized in that, In step 1, the sensor pose is acquired. and depth images Then, for depth images Each of them has pixel coordinates Depth value Perform the following operations on the pixels: in, for That is, the coordinates of a pixel in the world coordinate system. To obtain the current pose of the sensor, This is the camera's intrinsic parameter matrix; repeat the above steps for each depth image, converting each pixel into a point in the world coordinate system, and accumulating these point clouds to form the current moment. Complete 3D point cloud Next, using point cloud registration methods, a global map is incrementally constructed by utilizing the complete 3D point cloud at the current moment and the 3D point clouds at all past moments. 。 3. The UAV layered exploration method based on gravity mechanism viewpoint generation as described in claim 2, characterized in that, In step 4, whether all meet the narrow requirement is: abandoned boundary and Three-dimensional coordinates of any voxel Is the average distance between cluster positions less than a set threshold? 4. The UAV layered exploration method based on gravity mechanism viewpoint generation as described in claim 1, characterized in that, In step 5, the viewpoint generation method is determined based on whether the scene is detailed or wide. This includes: if it is a detailed scene, the viewpoint is generated using a narrow-channel gravity mechanism. These viewpoints are from sampling points Composition; sampling points Expressed in polar coordinates, with the origin representing the average position of the front edge. ： Among them, the polar diameter The distribution conforms to: polar angle The distribution conforms to: If sampling points If the viewpoint overlaps with an occupied or unknown voxel, it is resampled; if the scene is wide, the viewpoint is updated using an interactive lateral gravity mechanism. These viewpoints are from sampling points Composition; sampling points Expressed in polar coordinates, with the origin representing the average position of the front edge. ： Among them, the polar diameter The distribution conforms to: polar angle The distribution conforms to: in The current average position of the frontier Pointing to the drone's own coordinates and the current average position of the frontier Pointing to sampling point The angle between them; above All of these are fixed values.

5. The UAV layered exploration method based on gravity mechanism viewpoint generation as described in claim 4, characterized in that, In step 6, the ATSP cost matrix is ​​generated. The methods include: utilizing the current set Generate ATSP cost matrix , of which elements Meaning the first The forefront to the first The cost of being at the forefront, Meaning: The drone's location is up to the... The cost of being at the forefront, Meaning the first The cost from the front edge to the drone's location is generated as follows: the elements in the first column are: The elements in the first row are: in Indicates the number of viewpoints. For the drone's own coordinates, for The Middle A cutting-edge Average position of viewpoint , For the drone's velocity vector, Indicates the drone has reached the The path length of each frontier; in addition, the ATSP cost matrix Except for the first row and the first column, all others The elements of the square array are: 。 6. The UAV layered exploration method based on gravity mechanism viewpoint generation as described in claim 5, characterized in that, In step 6, the LKH solver is used to solve the matrix to obtain the frontier. The access sequence is used as the UAV waypoint sequence. And execute.

Images