A semantic perception RRT path planning method adaptive to texture degradation environment

Abstract

This invention discloses a semantically perceptive, rapidly expanding random tree (RRT) path planning method adapted to texture degradation environments. The method first acquires RGB and depth images for visual navigation. Then, it performs semantic segmentation on the input RGB images and assigns different weights to different semantic classes using a fuzzy evaluation method. An occupancy grid map is then combined with the semantic mask image to generate an information map for planning. Subsequently, heuristic sampling is performed on the information map, selecting nodes with higher information gain as nodes in the tree structure, and then connecting and expanding them. This method uses semantic information to guide the UAV to fly towards feature-rich scenes, solving the problem of visual positioning drift that occurs when the UAV flies to areas with degraded environmental textures during path search using the RRT method. Using the method proposed in this invention enables the UAV to reach the predetermined target point more safely and efficiently.

Description

Technical Field

[0001] This invention relates to a semantically perceptual fast exploration random tree path planning method adapted to texture degradation environments, belonging to the field of UAV path planning. Background Technology

[0002] The widespread application of unmanned aerial vehicles (UAVs) in fields such as industrial inspection, package delivery, environmental monitoring, agricultural spraying, and geographic mapping has attracted considerable attention. In these applications, UAVs need strong autonomous flight capabilities to achieve safe and efficient flight in complex, unknown, and dynamic environments. To achieve this goal, researchers and engineers have developed various path planning methods to solve the challenges of autonomous flight in different scenarios.

[0003] Among them, the Rapid Expanding Random Tree (RRT) method, a sampling-based path planning method, has been widely applied to autonomous flight of UAVs. Autonomous navigation is a crucial aspect of autonomous flight for UAVs. Typically, UAVs can use GPS for autonomous navigation in outdoor environments. However, GPS can fail in certain situations, such as under trees or between tall buildings. Vision-based navigation is an alternative to GPS. Currently, vision-based simultaneous localization and mapping (SLAM) is widely used for autonomous navigation of UAVs.

[0004] However, traditional Regression-Based Navigation (RRT) methods may fall short in navigation environments with degraded visual textures. For example, when a UAV flies in areas where visual textures are degraded or absent, vision-based navigation methods can be severely affected, leading to decreased accuracy in localization and mapping. This is because the paths planned by traditional RRT methods do not take into account the need for texture information in visual navigation, and the UAV may enter areas lacking texture, thus impacting its autonomous navigation performance.

[0005] Therefore, this invention proposes an improved RRT path planning method to avoid flying to areas with degraded or vanished textures, thereby improving the autonomous flight performance of UAVs in environments with degraded or vanished textures. This improved method, based on traditional RRT methods, utilizes environmental texture information contained in semantically segmented images, increasing the weight of texture information that is helpful for autonomous navigation to guide the UAV towards feature-rich scenes, thus improving the navigation accuracy and safety of UAVs in environments with uneven texture complexity. Summary of the Invention

[0006] The purpose of this invention is to propose a semantically perceptive, rapidly expanding random tree path planning method adapted to texture degradation environments. Utilizing RGB images and depth maps, an information map required for path planning is constructed. A novel heuristic sampling method is used to select sampling points and connect and expand them. Simultaneously, a rewiring mechanism-based method is designed to ensure the planned path contains as much information as possible while minimizing path distance, thus achieving the shortest possible flight path for the UAV while maintaining accurate positioning. Specific steps include:

[0007] Step 1: Input RGB image and depth image to create an information map.

[0008] In a non-open environment, using a binocular camera at a frequency of 15Hz, RGB images and depth images with a resolution of 1024×768 are acquired. The depth image is used to construct a 3D occupancy grid map of the drone's local environment for obstacle avoidance. The RGB images are semantically segmented to obtain semantic mask images, and a fuzzy evaluation method is used to evaluate and score each semantic class in the scene, assigning it different weight values. The aim is to guide the drone to areas with rich environmental texture information. The semantic classes are derived from the scalar class labels of the Cityscapes class, namely {buildings, roads, sky, vehicles, humans, nature}. Using optical flow and combining the semantic images, a 2D top-down projection map of the environment within the drone's field of view is generated, where each grid in the projection map contains semantic information. Specifically, the following sub-steps are included:

[0009] Sub-step 1: Determine the weights of different semantic classes

[0010] Establish a factor set U = (u1, u2, ..., u m ), where element u i Representing the i-th factor influencing visual localization, establish a comprehensive evaluation set V = (v1, v2, ..., v...). n ), where element v j This represents the j-th evaluation result.

[0011] Perform a single-factor fuzzy evaluation to obtain the fuzzy comprehensive evaluation matrix. The result of the single-factor evaluation of the i-th element is represented by a fuzzy set as: R i =(r i1 r i2 , ..., r in ), with m single-factor evaluation sets R1, R2, ..., R m The rows form a matrix R m×n This is called the fuzzy comprehensive evaluation matrix. The weights of each factor are determined using the entropy weight method, and the fuzzy set of the weights of each factor is denoted by A:

[0012] A = (a1, a2, ..., a...) m (12)

[0013] A comprehensive evaluation model is established. After determining the single-factor evaluation matrix R and the factor weight vector A, the fuzzy vector A on U is transformed into the fuzzy vector B on V through fuzzy transformation, i.e.

[0014]

[0015] In the formula This is called the comprehensive evaluation synthesis operator, which is taken here as a matrix multiplication operation.

[0016] Once the comprehensive evaluation model is determined, the system score is determined, i.e. Where F is the total score of the system, and S is the classification of the corresponding factor in V.

[0017] Sub-step 2: Create an information map

[0018] Within the drone's line of sight, the instantaneous overhead view is integrated into a sliding window semantic map centered on the drone, utilizing the drone's position. This map serves as a global map to guide the drone to areas rich in visual features. The definition of the information map is as follows:

[0019]

[0020] In the formula, N is the total number of grid cells in the information map. The center position P of each grid cell is... i =[P xi P yi Semantic class labels are derived from descriptor vector C. i Description. That is, C. i ∈{buildings, roads, sky, vehicles, humans, nature}.

[0021] Step 2: Utilize information maps to autonomously plan routes for the drone.

[0022] To prevent UAVs from flying into areas with degraded environmental textures during autonomous navigation, this invention employs an improved RRT method for path planning on an information map M. Compared to traditional sampling-based path planning methods, this invention proposes using semantic information to guide tree expansion. The specific steps of the method are as follows:

[0023] Sub-step 1): Heuristic sampling

[0024] This invention proposes a heuristic sampling method to obtain new sampling points, the specific steps of which are as follows:

[0025] (1) Generate n uniformly in the workspace using the function Random(·). s random sampling points x i ,Right now It is important to note that the number of sampling points has a significant impact on the efficiency of the method. If the number of sampling points is too large, it will take too much time to search for the optimal sampling points; if the number of sampling points is too small, the selected sampling points may not be optimal.

[0026] (2) The information gain of each sampling point is obtained using an information map M. A triangular mesh is introduced, the size of which depends on the camera's field of view. This triangle is divided uniformly. Only when the information contained in each small mesh is greater than the threshold ε (the minimum information value that ensures the UAV can perform visual positioning normally) will that mesh be included in the calculation. The information gain of each sampling point is defined as follows:

[0027] Info(x i )=||p(x i )-p(TgridCenter)||·Free(x i (15)

[0028] In the formula, p(·) represents the coordinates on the 2D projected map, and TgridCenter represents the coordinates of the center point of the uniformly divided small triangular grid. The meaning of the grid center point mentioned below is the same as that of TgridCenter. The triangular grid involved in the calculation contains the most information among all grids, and the grids closer to the UAV are given priority. Its definition is as follows:

[0029]

[0030] In the formula, Δd represents the distance between the grid center and the UAV, and the Normalization() function represents the normalization of the information score of the grid, which is defined as follows:

[0031]

[0032] In the formula F i The scores represent different semantic classes. Ngridi represents the number of rasters of the same semantic class. When counting semantic classes that do not completely fill a raster, a semantic class is counted as 1 if its area is greater than half of the entire raster, and 0 otherwise. The function Free(·) is defined using depth information, as follows:

[0033]

[0034] Heuristic sampling points The most insightful sampling points from these random sampling points are added to the tree structure.

[0035] Sub-step 2): Nearest neighbor

[0036] Use the function Nearest(·) to get the nearest data in the tree. Mid-range x info The nearest point x near x near =argmin x∈V ||x info -x|| In the formula, V represents the set of collision-free nodes, and E represents the set of edges connecting collision-free nodes.

[0037] Sub-step 3): Guide node

[0038] Get a new node x using the function Steer(·) new In x near Chao X info x is obtained in the direction of movement with a fixed step size Δ. new .

[0039] Sub-step 4): Rerouting mechanism

[0040] To generate more path information and shorter path length, this invention proposes a rewiring mechanism to modify the generated path. For a path consisting of points N = {n0, n1, ..., n...} j The total information gain obtained by the initial path ζ formed by} is calculated as follows:

[0041] Info(ζ)=∑Info(n i (19)

[0042] A new sampling point x will be obtained through the sampling strategy proposed in this invention. info Node x new Obtain and check x using the Steer() function. new The surrounding points, taking the points in the tree within a certain range as x new The set of neighboring nodes X neighbor For neighbor node x neighbor ∈X neighbor The path connecting it to the starting point is node N. visited ={n0, n2, n3, n4} and edge e i The network consists of nodes i ∈ I1 = {1, 2, 4}. The availability of neighboring nodes is defined as follows:

[0043]

[0044] In the formula, the function L(·) represents the length of each side.

[0045] If U1 > U2, it means there exists a better path that not only gains more information but is also shorter. In this case, node x... new The parent node is changed to x neighbor This concludes a rewiring process. Whenever a new node x... new Once generated, the rerouting process begins. The rerouting mechanism proposed in this invention can modify the path to one with more information and lower cost, which can greatly improve the efficiency of the method.

[0046] Sub-step 5): Repeat sub-steps 1) to 4) until the number of heuristic sampling points reaches a preset value.

[0047] Sub-step 6): Scrolling optimization

[0048] Pruning the tree structure preserves a path that is both information-rich and close to the target, defined as follows:

[0049]

[0050] In the formula, α is a weighting factor that prioritizes information gain, and the function Cost(·) represents the degree to which a path is close to the destination, as defined below:

[0051]

[0052] In the formula, Δdis represents the distance between the node at the end of the trajectory and the target point, and d cur d represents the distance between the drone's current position and the target point. attract This represents the region with strong attraction near the target point, and k represents the degree of dependence on the current distance.

[0053] The beneficial effects of this invention are as follows: the semantic perception fast exploration random tree path planning method proposed in this invention can utilize environmental texture information in the environment to avoid the UAV entering areas with scarce environmental texture; and a rewiring mechanism is designed to ensure that the UAV's flight path is rich in information while minimizing the flight path distance. Attached Figure Description

[0054] Figure 1 This is a flowchart of a method for rapidly expanding random trees based on semantic information;

[0055] Figure 2 It is a two-dimensional top-view projection;

[0056] Figure 3 This is a schematic diagram of the rewiring mechanism. Detailed Implementation

[0057] The widespread application of unmanned aerial vehicles (UAVs) in fields such as industrial inspection, package delivery, environmental monitoring, agricultural spraying, and geographic mapping has attracted considerable attention. In these applications, UAVs need strong autonomous flight capabilities to achieve safe and efficient flight in complex, unknown, and dynamic environments. To achieve this goal, researchers and engineers have developed various path planning methods to solve the challenges of autonomous flight in different scenarios.

[0058] Among them, the Rapid Expanding Random Tree (RRT) method, a sampling-based path planning method, has been widely applied to UAV autonomous flight. Autonomous navigation is another crucial aspect of UAV autonomous flight. Typically, UAVs can use GPS for autonomous navigation in outdoor environments. However, GPS can fail in certain situations, such as under trees or between tall buildings. Vision-based navigation is an alternative to GPS. Currently, vision-based simultaneous localization and mapping (SMR) is widely used for UAV autonomous navigation. However, in navigation environments with degraded visual textures, traditional RRT methods may be insufficient. For example, when a UAV flies in areas with degraded or absent visual textures, vision-based navigation methods can be severely affected, leading to decreased accuracy in localization and mapping. This is because the paths planned by traditional RRT methods do not consider the need for texture information in visual navigation; the UAV may enter areas lacking texture, thus impacting its autonomous navigation performance.

[0059] Therefore, this invention proposes an improved RRT path planning method to prevent UAVs from flying to areas with degraded or disappeared textures, thereby improving the autonomous flight performance of UAVs in environments with degraded or disappeared textures. This improved method, based on the traditional RRT method, utilizes environmental texture information contained in semantically segmented images, increasing the weight of texture information that is helpful for localization to guide the UAV towards feature-rich scenes, thus improving the navigation accuracy and safety of the UAV in environments with uneven texture complexity. This invention uses semantic information to guide the UAV towards feature-rich scenes, solving the problem of visual localization drift that occurs when the UAV flies to areas with degraded environmental textures during path search using the RRT method. Using the method proposed in this invention enables UAVs to reach predetermined target points more safely and efficiently.

[0060] The purpose of this invention is to propose a semantically perceptive, rapidly expanding random tree path planning method adapted to texture degradation environments. Utilizing RGB images and depth maps, an information map required for path planning is constructed. A novel heuristic sampling method is used to select sampling points and connect and expand them. Simultaneously, a rewiring mechanism-based method is designed to ensure the planned path contains as much information as possible while minimizing path distance, thus achieving the shortest possible flight path for the UAV while maintaining accurate positioning. Specific steps include:

[0061] Step 1: Input RGB image and depth image to create an information map.

[0062] In a non-open environment, using a binocular camera at a frequency of 15Hz, RGB images and depth images with a resolution of 1024×768 are acquired. The depth image is used to construct a 3D occupancy grid map of the drone's local environment for obstacle avoidance. The RGB images are semantically segmented to obtain semantic mask images, and a fuzzy evaluation method is used to evaluate and score each semantic class in the scene, assigning it different weight values. The aim is to guide the drone to areas with rich environmental texture information. The semantic classes are derived from the scalar class labels of the Cityscapes class, namely {buildings, roads, sky, vehicles, humans, nature}. Using optical flow and combining the semantic images, a 2D top-down projection map of the environment within the drone's field of view is generated, where each grid in the projection map contains semantic information. Specifically, the following sub-steps are included:

[0063] Sub-step 1: Determine the weights of different semantic classes

[0064] Establish a factor set U = (u1, u2, ..., u m ), where element u i Representing the i-th factor influencing visual localization, establish a comprehensive evaluation set V = (v1, v2, ..., v...). n ), where element v j This represents the j-th evaluation result.

[0065] Perform a single-factor fuzzy evaluation to obtain the fuzzy comprehensive evaluation matrix. The result of the single-factor evaluation of the i-th element is represented by a fuzzy set as: R i =(r i1 r i2 , ..., r in ), with m single-factor evaluation sets R1, R2, ..., R m The rows form a matrix R m×n This is called the fuzzy comprehensive evaluation matrix. The weights of each factor are determined using the entropy weight method, and the fuzzy set of the weights of each factor is denoted by A:

[0066] A = (a1, a2, ..., a...) m ) (twenty three)

[0067] A comprehensive evaluation model is established. After determining the single-factor evaluation matrix R and the factor weight vector A, the fuzzy vector A on U is transformed into the fuzzy vector B on V through fuzzy transformation, i.e.

[0068]

[0069] In the formula This is called the comprehensive evaluation synthesis operator, which is taken here as a matrix multiplication operation.

[0070] Once the comprehensive evaluation model is determined, the system score is determined, i.e. Where F is the total score of the system, and S is the classification of the corresponding factor in V.

[0071] Sub-step 2: Create an information map

[0072] Within the drone's line of sight, the instantaneous overhead view is integrated into a sliding window semantic map centered on the drone, utilizing the drone's position. This map serves as a global map to guide the drone to areas rich in visual features. The definition of the information map is as follows:

[0073]

[0074] In the formula, N is the total number of grid cells in the information map. The center position P of each grid cell is... i =[P xi P yi Semantic class labels are derived from descriptor vector C. i Description. That is, C. i ∈{buildings, roads, sky, vehicles, humans, nature}.

[0075] Step 2: Utilize information maps to autonomously plan routes for the drone.

[0076] To prevent UAVs from flying into areas with degraded environmental textures during autonomous navigation, this invention employs an improved RRT (Regression-Based Path Planning) method on an information map M for path planning. Compared to traditional sampling-based path planning methods, this invention proposes using semantic information to guide tree expansion. The flowchart of the method is as follows: Figure 1 As shown, the specific steps of the method are as follows:

[0077] Sub-step 1): Heuristic sampling

[0078] This invention proposes a heuristic sampling method to obtain new sampling points, the specific steps of which are as follows:

[0079] (1) Generate n uniformly in the workspace using the function Random(·). s random sampling points x i,Right now It is important to note that the number of sampling points has a significant impact on the efficiency of the method. If the number of sampling points is too large, it will take too much time to search for the optimal sampling points; if the number of sampling points is too small, the selected sampling points may not be optimal.

[0080] (2) The information gain of each sampling point is obtained using an information map M. A triangular mesh is introduced, the size of which depends on the camera's field of view. This triangle is divided uniformly. Only when the information contained in each small mesh is greater than the threshold ε (the minimum information value that ensures the UAV can perform normal visual positioning) will that mesh be included in the calculation. A schematic diagram is shown below. Figure 2 As shown in the figure. Different colors in the figure represent different semantic classes. The information gain of each sampling point is defined as follows:

[0081] Info(x i )=||p(x i )-p(TgridCenter)||·Free(x i (26)

[0082] In the formula, p(·) represents the coordinates on the 2D projected map, and TgridCenter represents the coordinates of the center point of the uniformly divided small triangular grid. The meaning of the grid center point mentioned below is the same as that of TgridCenter. The triangular grid involved in the calculation contains the most information among all grids, and the grids closer to the UAV are given priority. Its definition is as follows:

[0083]

[0084] In the formula, Δd represents the distance between the grid center and the UAV, and the Normalization() function represents the normalization of the information score of the grid, which is defined as follows:

[0085]

[0086] In the formula F i Ngrid represents scores for different semantic classes. i This represents the number of rasters with the same semantic class. When counting semantic classes that do not completely fill a raster, it is defined that if the area occupied by a semantic class is greater than half of the entire raster, it is counted as 1; otherwise, it is counted as 0. The function Free(·) is defined using depth information, as follows:

[0087]

[0088] Heuristic sampling points The most insightful sampling points from these random sampling points are added to the tree structure.

[0089] Sub-step 2): Nearest neighbor

[0090] Use the function Nearest(·) to get the nearest data in the tree. Mid-range x info The nearest point x near x near =argmin x∈V ||x info -x|| In the formula, V represents the set of collision-free nodes, and E represents the set of edges connecting collision-free nodes.

[0091] Sub-step 3): Guide node

[0092] Get a new node x using the function Steer(·) new In x nesr Chao X info x is obtained in the direction of movement with a fixed step size Δ. new .

[0093] Sub-step 4): Rerouting mechanism

[0094] To generate more path information and shorter path length, this invention proposes a rewiring mechanism to modify the generated path. For a path consisting of points N = {n0, n1, ..., n...} j The total information gain obtained by the initial path ζ formed by} is calculated as follows:

[0095] Info(ζ)=∑Info(n i (30)

[0096] like Figure 3 As shown, a new sampling point x will be obtained through the sampling strategy proposed in this invention. info Node x new Obtain and check x using the Steer() function. new The surrounding points, taking the points in the tree within a certain range as x new The set of neighboring nodes X neighbor For neighbor node x neighbor ∈X neighbor The path connecting it to the starting point is node N. visited ={n0, n2, n3, n4} and edge e i The network consists of nodes i ∈ I1 = {1, 2, 4). The availability of neighboring nodes is defined as follows:

[0097]

[0098] In the formula, the function L(·) represents the length of each side.

[0099] If U1 > U2, it means there exists a better path that not only gains more information but is also shorter. In this case, node x... new The parent node is changed to x neighbor This concludes a rewiring process. Whenever a new node x... new Once generated, the rerouting process begins. The rerouting mechanism proposed in this invention can modify the path to one with more information and lower cost, which can greatly improve the efficiency of the method.

[0100] Sub-step 5): Repeat sub-steps 1) to 4) until the number of heuristic sampling points reaches a preset value.

[0101] Sub-step 6): Scrolling optimization

[0102] Pruning the tree structure preserves a path that is both information-rich and close to the target, defined as follows:

[0103]

[0104] In the formula, α is a weighting factor that prioritizes information gain, and the function Cost(·) represents the degree to which a path is close to the destination, as defined below:

[0105]

[0106] In the formula, Δdis represents the distance between the node at the end of the trajectory and the target point, and d cur d represents the distance between the drone's current position and the target point. attract This represents the region with strong attraction near the target point, and k represents the degree of dependence on the current distance.

[0107] The semantic-aware fast exploration random tree path planning method proposed in this invention can utilize environmental texture information in the environment to ensure that the UAV avoids entering areas with scarce environmental texture; and a rewire mechanism is designed to ensure that the UAV's flight path is information-rich while minimizing the flight path distance.

Claims

1. A semantically aware RRT path planning method adapted to texture degradation environments, characterized in that, Using RGB images and depth maps, an information map is constructed for path planning. A heuristic sampling method is used to select sampling points, connect and expand them. Simultaneously, a rewiring mechanism-based method is designed to make the planned path contain more information and shorter distances, achieving the shortest flight path for the UAV while ensuring accurate positioning. Specific steps include: Step 1: Input RGB image and depth image to create an information map. In a non-open environment, using a binocular camera at a frequency of 15Hz, RGB images and depth images with a resolution of 1024×768 are acquired. The depth image is used to construct a 3D occupancy grid map of the drone's local environment for obstacle avoidance. The RGB images are semantically segmented to obtain semantic mask images, and a fuzzy evaluation method is used to evaluate and score each semantic class in the scene, assigning it different weight values. The purpose is to guide the drone to areas with rich environmental texture information. The semantic classes are derived from the scalar class labels of the Cityscapes class, namely {buildings, roads, sky, vehicles, humans, nature}. Using optical flow and combining the semantic images, a 2D top-down projection map of the environment within the drone's field of view is generated. Each grid in the projection map contains semantic information. The specific steps include the following: Sub-step 1: Determine the weights of different semantic classes Establish factor set , of which elements Represents the first factor affecting visual positioning Given several factors, i = {1, 2, ..., m}, establish an evaluation set for comprehensive evaluation. , of which elements Representing the The evaluation results are given by j={1,2,…,n}; Perform single-factor fuzzy evaluation to obtain the fuzzy comprehensive evaluation matrix; The results of a single-factor evaluation of an element are represented by a fuzzy set as follows: ,by Single-factor evaluation set Form a matrix by rows This is called the fuzzy comprehensive evaluation matrix; the weights of each factor are determined using the entropy weight method, and the fuzzy set of the weights of each factor is used... express: ； Establish a comprehensive evaluation model; determine the single-factor evaluation matrix. Sum of Factor Weights Then, through fuzzy transformation fuzzy vectors on Become fuzzy vectors on ,Right now ； In the formula This is called the comprehensive evaluation synthesis operator, which is represented here as matrix multiplication. Once the comprehensive evaluation model is determined, the system score is determined, i.e. in The total score for the system. for The classification of corresponding factors; Sub-step 2: Create an information map Within the drone's visual range, the instantaneous overhead view is integrated into a sliding window semantic map centered on the drone, utilizing the drone's position. This map serves as a global map to guide the drone to areas rich in visual features. The definition of the information map is as follows: ； In the formula, The total number of grid cells in the information map; the center position of each grid cell. Semantic class labels are composed of descriptor vectors Description; i.e. ; Step 2: Utilize information maps to autonomously plan routes for the drone. To prevent drones from flying into areas with degraded environmental textures during actual autonomous navigation, information maps are used. An improved RRT method is used for path planning. Compared with the traditional sampling-based path planning method, this method utilizes semantic information to guide tree expansion. The specific steps of the method are as follows: Sub-step 1): Heuristic sampling A heuristic sampling method is proposed to obtain new sampling points. The specific steps are as follows: (1) Through the function Uniform generation in the workspace Random sampling points ,Right now The number of sampling points has a significant impact on the efficiency of the method. If the number of sampling points is too large, too much time will be spent searching for the optimal sampling points. If the number of sampling points is too small, the selected sampling points may not be optimal. (2) Using information maps To obtain the information gain of each sampling point, a triangular mesh is introduced, the size of which depends on the camera's field of view. This triangle is uniformly divided, and only when the information contained in each small mesh is greater than a threshold is the information gain determined. At that time, the threshold The grid is only included in the calculation to ensure the minimum information value required for the drone to perform visual positioning correctly; the information gain of each sampling point is defined as follows: ； In the formula, Represents the coordinates on a 2D projected map. This represents the coordinates of the center point of the uniformly divided small triangular mesh; the triangular mesh used in the calculation contains the most information of all meshes, and meshes closer to the UAV are given priority, as defined below: ； In the formula This indicates the distance between the grid center and the drone. The function represents the normalization of the information score of the grid, and its definition is as follows: ； In the formula , Scores representing different semantic classes This represents the number of grid cells of the same semantic class. When counting semantic classes that do not completely fill a grid cell, it is defined that if the area occupied by a semantic class is greater than half of the entire grid cell, it is counted as 1; otherwise, it is counted as 0. Based on the definition of depth information, it is defined as follows: ； Heuristic sampling points The most insightful sampling points from these random sampling points are added to the tree structure; Sub-step 2): Nearest neighbor Using functions Get in tree Mid-range nearest point , In the formula Represents the set of collision-free nodes. The set of edges connecting collision-free nodes; Sub-step 3): Bootstrap node Through function Get a new node ;exist towards The movement proceeds in a fixed step size. To obtain ; Sub-step 4): Rewiring mechanism To generate more path information and shorter path lengths, a rewiring mechanism for modifying the generated path is proposed; for a path originating from a point... The initial path formed The total information gain obtained is calculated as follows: ； A new sampling point will be obtained through the proposed sampling strategy. ,node pass Function retrieval and inspection The surrounding points, taking the points in the tree within a certain range as... The set of neighboring nodes For neighboring nodes The path connecting it to the starting point is formed by nodes. and edge Composition; defining the availability of neighboring nodes: ； In the formula, the function Indicates the length of each side; if This means there exists a better path that not only gains more information but also has a shorter path; at this point, the node... The parent node is changed to This concludes a rewiring process; whenever a new node... Once generated, the rerouting process will begin; the proposed rerouting mechanism can modify the path to a path with more information and lower path cost, which can improve the efficiency of the method. Sub-step 5): Repeat sub-steps 1) to 4) until the number of heuristic sampling points reaches a preset value; Sub-step 6): Scrolling optimization Pruning the tree structure preserves a path that is both information-rich and close to the target, defined as follows: ； In the formula It is a weighting factor that prioritizes information gain, the function This indicates how close a path is to its destination, and is defined as follows: ； In the formula, This represents the distance between the node at the end of the trajectory and the target point. This indicates the distance between the drone's current position and the target point. This indicates an area with strong attraction near the target point. It indicates the degree of dependence on the current distance.

Images