A method for unmanned aerial vehicle cluster distribution planning based on deep learning

Abstract

The application discloses a kind of based on deep learning's unmanned aerial vehicle cluster distribution planning method, comprising: according to the environment where flight task is located to construct digital matrix map, determines start point and end point coordinate;The trajectory position point of cluster geometric center is dynamically planned using the maximum gap path planning algorithm based on A-Star algorithm;Image of the surrounding area of trajectory position point is obtained, after processing, input is used for planning the shape of cluster deep neural network model, and the shape information of cluster planned is output;Then, according to the shape information of cluster and the surrounding environment of trajectory point, the input data group of cluster shape deflection angle model is obtained, and the deflection angle of cluster at trajectory point is output, so as to determine the distribution of cluster real-time environment self-adaptation.The application constructs the end-to-end "trajectory-shape-direction" distribution planning mechanism, significantly improves the rationality of unmanned aerial vehicle cluster distribution under dynamic environment and planning efficiency, and provides technical support for intelligent and autonomous operation of cluster.

Description

Technical Field

[0001] This invention belongs to the field of UAV swarm path planning technology, and in particular relates to a UAV swarm path planning method for complex non-uniform obstacle areas. Background Technology

[0002] With the development of artificial intelligence technology and unmanned aerial systems, drone swarms have demonstrated enormous potential in numerous application scenarios due to their strong collaborative operation capabilities, wide coverage, and high mission efficiency. Especially in tasks such as disaster detection, target search, airspace patrol, environmental monitoring, and mass incident response, drone swarms can accomplish complex tasks that a single drone cannot perform independently through multi-drone collaboration. Therefore, how to achieve intelligent and dynamic distribution planning for drone swarms has become one of the current research hotspots in the field of unmanned systems.

[0003] During drone swarm flight, adjusting the spatial distribution of each drone in real time according to environmental changes to ensure the swarm can efficiently traverse obstacle areas and maintain a reasonable geometric shape is a key technology for improving the success rate and safety of the system. Existing swarm distribution planning methods mainly rely on manually designed rule systems or heuristic algorithms, such as those based on formation maintenance, artificial potential fields, or particle swarm optimization. While these methods are effective for certain tasks, they are difficult to adapt to complex dynamic environments, especially in situations with dense obstacles and varied tasks, often exhibiting problems such as response lag, suboptimal paths, and susceptibility to local optima.

[0004] In recent years, deep learning technology has achieved remarkable results in image recognition, path prediction, and autonomous decision-making, providing new ideas for intelligent planning of UAV swarms. By training deep neural networks, they can autonomously learn the implicit relationships between environmental features and task requirements, thereby outputting swarm distribution strategies with global optimization characteristics. This holds promise for overcoming the bottlenecks of traditional methods in terms of model generalization and decision-making efficiency. Especially when dealing with complex spatial structures and dynamic targets, deep learning methods can achieve intelligent reconstruction of UAV swarm structures and path adaptation through an end-to-end perception-decision integrated framework.

[0005] However, existing research on deep learning-based UAV path planning largely focuses on single-UAV scenarios, failing to adequately consider the spatial coordination and information sharing among multiple UAVs in a swarm. Furthermore, significant challenges remain in swarm shape adaptive modeling, multi-source environmental information fusion processing, and real-time network inference capabilities, limiting its effectiveness in real-world applications. Therefore, there is an urgent need to propose a deep learning-based UAV swarm distribution planning method that balances global environmental understanding with local dynamic adjustment capabilities, enabling adaptive adjustment of the swarm structure in complex environments and improving overall mission completion efficiency and flight safety. Summary of the Invention

[0006] To address the above technical problems, this invention provides a method for drone swarm distribution planning based on deep learning.

[0007] The technical solution adopted by this invention to solve its technical problem is:

[0008] A method for drone swarm distribution planning based on deep learning includes the following steps:

[0009] S1. After standardizing the geospatial data of the flight mission environment, the data is divided into grid cells to construct a digital matrix map. Passable areas and obstacle areas are marked in the digital matrix map. Then, the starting and ending coordinates of the UAV swarm mission are determined in the digital matrix map.

[0010] S2 utilizes a digital matrix map to obtain a gap cost field based on the distance between passable areas and obstacle areas. Then, based on the gap cost field, the maximum gap path planning algorithm of the A-Star algorithm is used to ensure that the path is far away from obstacles while satisfying feasibility, thereby planning the path and extracting the trajectory point of the geometric center of the UAV cluster from the path.

[0011] S3. Obtain the RGB top-view image of the area surrounding the trajectory point and perform preprocessing: convert the RGB top-view image into the corresponding uniform color according to the passable area and obstacle area in the digital matrix map; then input the deep neural network model used to plan the swarm shape of the drones, and output the planned swarm shape information.

[0012] S4. Based on the cluster shape information and the surrounding environment of the trajectory points, the input data set for the cluster shape deflection angle model is obtained. After inputting the model, the deflection angle of the cluster at the trajectory points is obtained, thereby determining the real-time adaptive distribution of the cluster in the environment.

[0013] Furthermore, step S1 includes:

[0014] S101, acquire geospatial data of the flight mission environment, including remote sensing images, building distribution information, and no-fly zone information;

[0015] S102, standardize and rasterize the geospatial data, convert the original spatial data into a two-dimensional spatial expression under a unified coordinate system, divide the flight area into raster units according to a preset resolution, generate a digital matrix map, and then mark the corresponding raster units as passable areas or obstacle areas in the digital matrix map with different markers.

[0016] S103, based on flight mission requirements and geospatial constraints, determines the start and end points of the mission.

[0017] S104: Map the starting and ending points to the digital matrix map and obtain the corresponding index coordinates, i.e., the grid cell number, in the digital matrix map.

[0018] Furthermore, step S2 includes:

[0019] S201, using the digital matrix map in S102, construct the gap cost field: based on the passable area, calculate the Euclidean distance value from each passable area grid cell to the nearest obstacle area grid cell through the distance transformation method to use as the minimum gap value, thereby obtaining the global obstacle distance field, and using the global obstacle distance field as the gap cost field to quantify the safety margin between each point and surrounding obstacles in path planning.

[0020] S202, Perform gap-guided A-Star path search: Based on the constructed gap cost field, define the path cost function as the minimum gap value of the current path node's location, and integrate the heuristic function to simultaneously consider the Euclidean distance from the current node to the target node and the reverse guidance effect of the gap value. By maximizing the minimum gap value of the path, ensure that the path is as far away from the obstacle edge as possible while satisfying feasibility, thereby forming an improved A-Star search strategy to plan the path.

[0021] S203 After the path planning is completed, a trajectory point is extracted from the trajectory points on the path at predetermined intervals, and used as a dynamic trajectory point set for the geometric center of the UAV swarm, forming a time-ordered reference trajectory for the geometric center of the swarm.

[0022] Furthermore, in step S202, the expression for the fusion heuristic function f(n) is: , where the path cost function , representing the minimum gap value for each point p on the path from the starting point s to node n. This represents the minimum clearance value from the starting point s to the path point p, while the maximum clearance value refers to the maximum safe distance between the drone and surrounding obstacles during its movement. It is a heuristic function of the A-Star algorithm, used to estimate the expected cost from the current node n to the target node.

[0023] Furthermore, step S3 includes:

[0024] S301, Collect RGB top-view images of the area surrounding each dynamic trajectory point in step S203;

[0025] S302, perform image preprocessing on the RGB top view image, including the following sub-steps:

[0026] 1) Scale the image to the preset input resolution size;

[0027] 2) Draw a colored square with a side length of a preset number of pixels at the center of the image to mark the location of the geometric center of the cluster;

[0028] 3) Re-encode the original environmental areas in the image according to the passable or obstacle areas marked by the grid cells: convert the passable and obstacle areas to different colors, and both are different from the color of the geometric center square of the cluster;

[0029] S303, the preprocessed image is input into the cluster shape recognition model S-DNN, which is a trained deep neural network model; after processing the image, the S-DNN obtains a multi-dimensional vector, where the dimension represents the preset number of formations, each value in the vector represents the probability of selecting a certain formation, and the formation corresponding to the largest vector value is taken as the planned cluster shape.

[0030] Furthermore, step S4 includes:

[0031] S401, based on the planned cluster shape output by S-DNN, using the forward direction from the current dynamic trajectory point of the planned cluster distribution to the next cluster trajectory point as the reference angle, generate an image of the cluster shape in the preprocessed RGB top view image, and simultaneously generate two images of the cluster shape in the RGB top view image rotated by a preset angle in opposite directions based on the reference angle, and use these three images as the input data set of the cluster deflection angle planning model O-DNN, wherein O-DNN is a trained deep neural network model;

[0032] The S402 O-DNN model outputs a single value ranging from -90° to 90°, representing the rotation angle of the drone swarm relative to the reference angle, which, together with the output of the S-DNN, determines the distribution information of the swarm.

[0033] S403 calculates the distribution location of each drone in the cluster based on the cluster's distribution information, thereby completing the distribution planning of the entire cluster.

[0034] Furthermore, in step S303, the deep neural network model S-DNN is trained through the following steps:

[0035] ① Construct a dataset for planning the distribution shape of the cluster. The dataset includes multiple RGB images. The RGB images represent the surrounding environment of the cluster location. The center of the image is a colored square with a side length of a preset number of pixels, which represents the location of the geometric center of the cluster. Different colors are used in the image to represent the passable area and the obstacle area respectively, and the colors of the square, the passable area and the obstacle area are consistent with the corresponding colors in S302.

[0036] ② Construct a deep neural network model S-DNN for planning the shape of the cluster; the network structure of the S-DNN includes convolutional layers, fully connected layers and an output layer; the convolutional layer includes multiple concatenated convolutional kernels, and there is a max pooling layer and a batch normalization layer between every two concatenated convolutional kernels; the feature map finally generated by the convolutional layer is flattened into a one-dimensional vector and input into the fully connected layer, and after being processed by the fully connected layer, it reaches the output layer as the output;

[0037] ③ Divide the dataset from step ① into a training set and a test set according to a predetermined ratio, and use them to train and test the S-DNN. Use the cross-entropy loss function and SGD optimizer to update parameters during model training.

[0038] Furthermore, in step S401, the deep neural network model O-DNN is trained through the following steps:

[0039] 1) Construct a dataset for planning the deflection angle of the cluster distribution shape. Each set of data in this dataset consists of 3 RGB images of cluster shapes with different rotation angles. The cluster shape is represented by a regular geometric shape with color at the center of the image. The azimuth angle of the cluster shape on one image is set as the reference angle from the forward direction from the current dynamic trajectory point of the cluster distribution being planned to the next cluster trajectory point. The cluster azimuth angles on the other two images are obtained by rotating the reference angle in opposite directions by a preset angle.

[0040] 2) Construct a deep neural network model O-DNN for planning the deflection angle of the cluster shape. The network structure of O-DNN is similar to that of S-DNN, including convolutional layers, fully connected layers, and an output layer. The convolutional layer includes multiple concatenated convolutional kernels, and there is a max pooling layer and a batch normalization layer between every two concatenated convolutional kernels. The feature map generated by the convolutional layer is flattened into a one-dimensional vector and then input into the fully connected layer. After being processed by the fully connected layer, it reaches the output layer as the output.

[0041] 3) Divide the dataset from step 1) into training and testing sets according to a predetermined ratio, and use them to train and test the O-DNN. The cross-entropy loss function and SGD optimizer are used to update parameters during model training.

[0042] The technical advantages of this invention are as follows: By proposing a maximum gap path planning algorithm for the geometric center trajectory of a swarm, the safety and path feasibility of the swarm traversing complex obstacle environments are effectively improved; a swarm shape recognition model S-DNN based on RGB image perception is introduced, realizing joint perception of environmental structure and task requirements, and automatically generating an appropriate swarm shape; a swarm deflection angle planning model O-DNN is designed, which completes intelligent adjustment of the swarm orientation through multi-angle image input, improving the overall environmental adaptability and spatial coordination capability of the swarm; the entire method constructs an end-to-end "trajectory-shape-direction" distribution planning mechanism, which significantly improves the distribution rationality and planning efficiency of UAV swarms in dynamic environments, providing technical support for the intelligent and autonomous operation of swarms. Attached Figure Description

[0043] Figure 1 This is a flowchart of an embodiment of the present invention.

[0044] Figure 2 The initial map for this embodiment of the invention includes a starting point, an ending point, and obstacles, wherein the red star at the top represents the starting point, the blue star at the bottom represents the ending point, and the black blocks represent obstacles.

[0045] Figure 3 The location of the trajectory point on the map for the geometric center of the UAV cluster planned in this embodiment of the invention.

[0046] Figure 4 The structures of the two deep neural network models S-DNN and O-DNN used for planning cluster distribution in this embodiment of the invention are shown.

[0047] Figure 5 This is the input data for planning the cluster shape model in an embodiment of the present invention.

[0048] Figure 6 This is the input data for planning the cluster shape deflection angle model in an embodiment of the present invention.

[0049] Figure 7 This embodiment of the invention generates a cluster distribution from the starting point to the ending point for a given map. Detailed Implementation

[0050] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0051] In one embodiment, such as Figure 1 As shown, a method for drone swarm distribution planning based on deep learning is provided, which includes the following steps:

[0052] S100: Construct a dataset for swarm distribution planning, and design and train a deep neural network model for swarm distribution planning based on the dataset. The deep neural network model includes a swarm shape recognition model (S-DNN) for planning the shape of the drone swarm, and a swarm deflection angle planning model (O-DNN) for planning the deflection angle of the drone swarm.

[0053] S200: Construct a digital matrix map based on the environment of the flight mission to determine the starting and ending coordinates of the UAV swarm mission. After standardizing the geospatial data of the flight mission environment, it is divided into grid cells to construct a digital matrix map. Passable areas and obstacle areas are marked on the digital matrix map, and then the starting and ending coordinates of the UAV swarm mission are determined from the digital matrix map.

[0054] S300: The maximum gap path planning algorithm based on the A-Star algorithm is used to dynamically plan the trajectory location of the geometric center of the swarm. By constructing a digital matrix map, the gap cost field is obtained based on the distance between the passable area and the obstacle area. Then, based on the gap cost field, the maximum gap path planning algorithm of the A-Star algorithm is used to ensure that the path is feasible while staying away from the obstacle, thereby planning the path and extracting the trajectory point of the geometric center of the UAV swarm from the path.

[0055] S400: Acquires RGB top-view images of the area surrounding the trajectory location point, performs image processing, and inputs them into an S-DNN to obtain the planned cluster shape information. First, it acquires RGB top-view images of the area surrounding the trajectory point and performs preprocessing: based on the passable area and obstacle area in the digital matrix map, the RGB top-view images are converted into corresponding uniform colors; then, it is input into a deep neural network model used to plan the cluster shape of the UAVs, and outputs the planned cluster shape information.

[0056] S500: Based on the shape information of the cluster and the surrounding environment of the trajectory points, the input data group of the O-DNN is obtained. The O-DNN output obtains the deflection angle of the cluster at the trajectory points, thereby determining the adaptive distribution of the cluster in real time.

[0057] In one embodiment, S100 includes:

[0058] S110: Construct a dataset for planning the distribution shape of the cluster. In this dataset of this embodiment, the image used is an RGB image with a pixel size of 128×128. The image represents the surrounding environment of the cluster location. The center of the image is a red square with a pixel side length of 5, which represents the location of the geometric center of the cluster. The black area of ​​the image represents the passable area in the environment, and the green area represents the obstacle area in the environment.

[0059] S120: A dataset for planning the deflection angles of the cluster distribution shape is then constructed. In this embodiment, each dataset consists of three RGB images of cluster shapes with different rotation angles, enabling the neural network model to better perceive the differences in cluster rotational motion. The image center represents the red, regular geometric shape of the cluster. The azimuth angle of the cluster shape in the second image is set as the forward direction from the current dynamic trajectory point of the planned cluster distribution to the next cluster trajectory point, and serves as the reference angle. The cluster azimuth angles in the first and third images are obtained by rotating the reference angle clockwise and counterclockwise by 15°, respectively.

[0060] S130: See also Figure 4 This embodiment designs a deep neural network model for planning cluster shapes, named the Cluster Shape Recognition Model (S-DNN). The input to the S-DNN is a 128×128 RGB image. The network structure includes convolutional layers, fully connected layers, and an output layer. The first and second convolutional kernels are 5×5, and the third and fourth are 4×4. Each convolutional kernel is followed by a 2×2 max-pooling layer with a stride of 2, and a batch normalization layer. The feature maps generated by the convolutional layers are then flattened into one-dimensional vectors and input to the fully connected layers. Finally, the output layer outputs the results. The output of the S-DNN is a multi-dimensional vector, where the dimension represents the number of preset cluster shapes (formations). Each value in the vector represents the probability of selecting a certain formation. The formation corresponding to the vector value with the highest probability is selected as the planned cluster shape. The dataset in step S110 is divided into a training set and a test set in a 9:1 ratio and used to train and test the S-DNN. During the training process, the parameters are updated using the cross-entropy loss function and the SGD optimizer until training is complete.

[0061] S140: Design a deep neural network model for planning the deflection angle of the drone swarm shape. In this embodiment, this model is named the Swarm Deflection Angle Planning Model, or O-DNN. In this embodiment, except for the input, output, and the number of nodes in the fully connected layers, the overall structure of O-DNN is the same as that of S-DNN. The input of O-DNN is three RGB images of the swarm shape with different rotation angles. The output of O-DNN is a single value between -90° and 90°, representing the rotation angle of the drone swarm relative to the reference angle. The dataset from step S120 is divided into training and testing sets in a 9:1 ratio and used to train and test O-DNN. During training, the parameters are updated using the cross-entropy loss function and the SGD optimizer until training is complete.

[0062] In one embodiment, S200 includes:

[0063] S210: Acquire geospatial data of the flight mission environment. This step includes collecting remote sensing images of the target area, building distribution information, and no-fly zone information. In actual execution, this specific data can be acquired through satellite remote sensing, geographic information systems, 3D laser scanning, or other known methods.

[0064] S220: The geospatial data is standardized and rasterized, the original spatial data is converted into a two-dimensional spatial expression under a unified coordinate system, and the flight area is divided into regular raster units according to a preset resolution to generate a digital matrix map. In this embodiment, the number 1 in the digital matrix map represents the raster unit as an obstacle area, and the number 0 represents the raster unit as a feasible area.

[0065] S230: Based on flight mission requirements and geospatial constraints, determine the coordinates of the mission's start and end points. This step includes selecting the start and end points within the flyable area from a digital matrix map and setting the coordinates according to the nature of the mission, such as surveillance, reconnaissance, transport, or search. The selected start and end points should meet communication, takeoff and landing, and safety boundary conditions.

[0066] S240: Map the starting point and ending point coordinates to the corresponding index positions in the digital matrix map. By finding the grid number corresponding to the starting point and ending point, obtain their index coordinates in the digital matrix for subsequent path search algorithm calls.

[0067] In one embodiment, S300 includes:

[0068] S310: Map information preprocessing, see [link / reference] Figure 2 The system acquires environmental map images and performs uniform size standardization processing, converting the original images into grayscale images. Through image binarization, it extracts passable areas and obstacle areas, where passable areas are represented by pixel values ​​of 255 and obstacle areas are represented by values ​​other than 255, thereby constructing a rasterized digital map model for path planning.

[0069] S320: Construct a gap cost field. Based on the passable area, calculate the Euclidean distance from each passable grid point to the nearest obstacle using a distance transformation method to obtain a global obstacle distance field. This distance field can be regarded as a gap cost field, used to quantify the safety margin between each point and surrounding obstacles in path planning. This makes path search more inclined to choose areas with larger obstacle gaps, improving path safety and cluster passability.

[0070] S330: See also Figure 3This paper proposes an improved A-Star path search strategy based on gap guidance. Building upon the constructed gap cost field, the path cost function is defined as the minimum gap value at the current path node's location. A heuristic function is then integrated to form this improved A-Star search strategy. In this embodiment, the heuristic function considers both the Euclidean distance from the current node to the target node and the reverse guidance effect of the gap value. By maximizing the minimum gap value, it ensures that the path stays as far away from obstacle edges as possible while meeting feasibility requirements, avoiding narrow areas and improving path robustness. The expression for the integrated heuristic function f(n) is as follows: , where the path cost function This represents the minimum gap value for each point p on the path from the starting point s to node n. It is a heuristic function of the A-Star algorithm, used to estimate the expected cost from the current node n to the target node.

[0071] S340: After path planning is completed, the obtained path is sampled, and key nodes in the path are extracted as a set of dynamic trajectory points, forming a time-ordered cluster geometric center reference dynamic trajectory point. This trajectory can be used as a basic reference for subsequent attitude adjustment, cooperative obstacle avoidance, and formation control of multiple UAV clusters, realizing cluster cooperative behavior with a high safety margin.

[0072] In one embodiment, S400 includes:

[0073] S410: Collect an RGB top-view image of the area surrounding the dynamic trajectory point of the cluster geometric center planned in step S340.

[0074] S420: See also Figure 5 Image preprocessing of the RGB top-view image includes the following sub-steps:

[0075] (1) Scale the image to the preset input resolution size of 128×128.

[0076] (2) Draw a red square with a side length of 5 pixels at the center of the image to mark the location of the geometric center of the cluster.

[0077] (3) Re-encode the original environmental areas in the image according to their accessibility: convert the accessible areas to black pixel values ​​and the obstacle areas to green pixel values.

[0078] (4) The final image contains only black and green areas except for the red square, in order to enhance the clarity of semantic expression and facilitate feature extraction by neural networks.

[0079] S430: The preprocessed image is input into the deep neural network model S-DNN. After training, S-DNN can identify high-level semantic features such as the geometry, boundary contours, and orientation of potential obstacles in the image. The input of S-DNN is the preprocessed RGB image, and the output is the classification information of the cluster target shape. Specifically, S-DNN processes the image to obtain a multi-dimensional vector, where the dimension represents the number of preset formations, and each value in the vector represents the probability of selecting a certain formation. The formation corresponding to the largest vector value is taken as the planned cluster shape.

[0080] In one embodiment, S500 includes:

[0081] S510: See also Figure 6 Based on the cluster shape information planned by the S-DNN output, three RGB images of the corresponding planned cluster shapes with different rotation angles are generated as input data for the O-DNN model. In this embodiment, the cluster shape azimuth angle in the second image is set as the reference angle from the forward direction from the current dynamic trajectory point of the planned cluster distribution to the next cluster trajectory point. The cluster azimuth angles in the first and third images are obtained by rotating the reference angle clockwise and counterclockwise by 15°, respectively.

[0082] S520: Plan the shape deflection angle of the cluster at the current trajectory point using an O-DNN model.

[0083] S530: See also Figure 7 Once all cluster distribution information is obtained, the distribution location of each drone in the cluster can be calculated, thereby completing the distribution planning of the entire cluster.

[0084] The foregoing has provided a detailed description of a deep learning-based UAV swarm distribution planning method provided by this invention. Specific embodiments have illustrated the principles and implementation methods of this invention, aiming to aid in understanding its core concepts. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the basic principles of this invention, and all such modifications and substitutions should be covered within the scope of protection of this invention.

Claims

1. A method for distributed planning of unmanned aerial vehicle (UAV) swarms based on deep learning, characterized in that, Includes the following steps: S1. After standardizing the geospatial data of the flight mission environment, the data is divided into grid cells to construct a digital matrix map. Passable areas and obstacle areas are marked in the digital matrix map. Then, the starting and ending coordinates of the UAV swarm mission are determined in the digital matrix map. S2 utilizes a digital matrix map to obtain a gap cost field based on the distance between passable areas and obstacle areas. Then, based on the gap cost field, the maximum gap path planning algorithm of the A-Star algorithm is used to ensure that the path is far away from obstacles while satisfying feasibility, thereby planning the path and extracting the trajectory point of the geometric center of the UAV cluster from the path. S3. Obtain the RGB top-view image of the area surrounding the trajectory point and perform preprocessing: convert the RGB top-view image into the corresponding uniform color according to the passable area and obstacle area in the digital matrix map; then input the deep neural network model used to plan the swarm shape of the drones, and output the planned swarm shape information. S4. Based on the cluster shape information and the surrounding environment of the trajectory points, the input data set for the cluster shape deflection angle model is obtained. After inputting the model, the deflection angle of the cluster at the trajectory points is obtained, thereby determining the real-time adaptive distribution of the cluster in the environment. Step S2 includes: S201, using the digital matrix map in S1, constructs the gap cost field: based on the passable area, the Euclidean distance value from each passable area grid cell to the nearest obstacle area grid cell is calculated using the distance transformation method as the minimum gap value, thereby obtaining the global obstacle distance field, and using the global obstacle distance field as the gap cost field to quantify the safety margin between each point and surrounding obstacles in path planning. S202, Perform gap-guided A-Star path search: Based on the constructed gap cost field, define the path cost function as the minimum gap value of the current path node's location, and integrate the heuristic function to simultaneously consider the Euclidean distance from the current node to the target node and the reverse guidance effect of the gap value. By maximizing the minimum gap value of the path, ensure that the path is as far away from the obstacle edge as possible while satisfying feasibility, thereby forming an improved A-Star search strategy to plan the path. S203 After the path planning is completed, a trajectory point is extracted from the trajectory points on the path at predetermined intervals, and used as a dynamic trajectory point set of the geometric center of the UAV swarm, forming a time-ordered reference trajectory for the geometric center of the swarm. Step S3 includes: S301, Collect RGB top-view images of the area surrounding each dynamic trajectory point in step S203; S302, perform image preprocessing on the RGB top view image, including the following sub-steps: 1) Scale the image to the preset input resolution size; 2) Draw a colored square with a side length of a preset number of pixels at the center of the image to mark the location of the geometric center of the cluster; 3) Re-encode the original environmental areas in the image according to the passable or obstacle areas marked by the grid cells: convert the passable and obstacle areas to different colors, and both are different from the color of the geometric center square of the cluster; S303, the preprocessed image is input into the cluster shape recognition model S-DNN, which is a trained deep neural network model; after processing the image, the S-DNN obtains a multi-dimensional vector, where the dimension represents the preset number of formations, each value in the vector represents the probability of selecting a certain formation, and the formation corresponding to the largest vector value is taken as the planned cluster shape. Step S4 includes: S401, based on the planned cluster shape output by S-DNN, using the forward direction from the current dynamic trajectory point of the planned cluster distribution to the next cluster trajectory point as the reference angle, generate an image of the cluster shape in the preprocessed RGB top view image, and simultaneously generate two images of the cluster shape in the RGB top view image rotated by a preset angle in opposite directions based on the reference angle, and use these three images as the input data set of the cluster deflection angle planning model O-DNN, wherein O-DNN is a trained deep neural network model; S402, the O-DNN model outputs a single value in the range of -90° to 90°, representing the rotation angle of the drone swarm relative to the reference angle, which, together with the output of the S-DNN, determines the distribution information of the swarm; S403 calculates the distribution location of each drone in the cluster based on the cluster's distribution information, thereby completing the distribution planning of the entire cluster.

2. The method according to claim 1, characterized in that, Step S1 includes: S101, acquire geospatial data of the flight mission environment, including remote sensing images, building distribution information, and no-fly zone information; S102, standardize and rasterize the geospatial data, convert the original spatial data into a two-dimensional spatial expression under a unified coordinate system, divide the flight area into raster units according to a preset resolution, generate a digital matrix map, and then mark the corresponding raster units as passable areas or obstacle areas in the digital matrix map with different markers. S103, based on flight mission requirements and geospatial constraints, determines the start and end points of the mission. S104: Map the starting and ending points to the digital matrix map and obtain the corresponding index coordinates, i.e., the grid cell number, in the digital matrix map.

3. The method according to claim 1, characterized in that, In step S202, the heuristic function is fused. f ( n The expression for ) is f ( n )=- g ( n )+ h ( n ), where the path cost function , representing the minimum gap value for each point p on the path from the starting point s to node n. dist_map [ p The minimum clearance value represents the distance from the starting point s to the path point p. The maximum clearance value refers to the maximum safe distance between the drone and surrounding obstacles during its movement. h ( n ) is a heuristic function of the A-Star algorithm, used to estimate the expected cost from the current node n to the target node.

4. The method according to claim 1, characterized in that, In step S303, the deep neural network model S-DNN is trained through the following steps: ① Construct a dataset for planning the distribution shape of the cluster. The dataset includes multiple RGB images. The RGB images represent the surrounding environment of the cluster location. The center of the image is a colored square with a side length of a preset number of pixels, which represents the location of the geometric center of the cluster. Different colors are used in the image to represent the passable area and the obstacle area respectively, and the colors of the square, the passable area and the obstacle area are consistent with the corresponding colors in S302. ② Construct a deep neural network model S-DNN for planning the shape of the cluster; the network structure of the S-DNN includes convolutional layers, fully connected layers and an output layer; the convolutional layer includes multiple concatenated convolutional kernels, and there is a max pooling layer and a batch normalization layer between every two concatenated convolutional kernels; the feature map finally generated by the convolutional layer is flattened into a one-dimensional vector and input into the fully connected layer, and after being processed by the fully connected layer, it reaches the output layer as the output; ③ Divide the dataset from step ① into a training set and a test set according to a predetermined ratio, and use them to train and test the S-DNN. Use the cross-entropy loss function and SGD optimizer to update parameters during model training.

5. The method according to claim 4, characterized in that, In step S401, the deep neural network model O-DNN is trained through the following steps: 1) Construct a dataset for planning the deflection angle of the cluster distribution shape. Each set of data in this dataset consists of 3 RGB images of cluster shapes with different rotation angles. The cluster shape is represented by a regular geometric shape with color at the center of the image. The azimuth angle of the cluster shape on one image is set as the reference angle from the forward direction from the current dynamic trajectory point of the cluster distribution being planned to the next cluster trajectory point. The cluster azimuth angles on the other two images are obtained by rotating the reference angle in opposite directions by a preset angle. 2) Construct a deep neural network model O-DNN for planning the deflection angle of the cluster shape. The network structure of O-DNN is similar to that of S-DNN, including convolutional layers, fully connected layers, and an output layer. The convolutional layer includes multiple concatenated convolutional kernels, and there is a max pooling layer and a batch normalization layer between every two concatenated convolutional kernels. The feature map generated by the convolutional layer is flattened into a one-dimensional vector and then input into the fully connected layer. After being processed by the fully connected layer, it reaches the output layer as the output. 3) Divide the dataset from step 1) into training and testing sets according to a predetermined ratio, and use them to train and test the O-DNN. The cross-entropy loss function and SGD optimizer are used to update parameters during model training.

Images