Autonomous trajectory planning and image acquisition method and system for UAV

Abstract

The present invention provides a method and system for autonomous trajectory planning and image acquisition of unmanned aerial vehicles, which relates to the field of unmanned aerial vehicle inspection technology. First, through intelligent three-dimensional grid division and dynamic route planning, the coverage and efficiency of feature area monitoring are significantly improved; secondly, the integration of real-time meteorological data and historical feature data enhances the pertinence and adaptability of acquisition, which helps to obtain high-quality images in complex and changeable environments; then, the combination of real-time processing and ground deep learning analysis greatly improves the accuracy and efficiency of feature extraction; based on the adaptive supplementary sampling mechanism of key focus areas, rapid response and precise positioning of suspected feature areas are achieved, providing reliable data support for subsequent detailed analysis and decision-making. The present invention is not only suitable for various recognition tasks of preset features, but also has good scalability and flexibility, and can be adjusted and optimized according to different application requirements.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle (UAV) inspection technology, and in particular to a method and system for autonomous trajectory planning and image acquisition of a UAV. Background Art

[0002] The current use of drones for image acquisition and path planning is an innovative inspection technology. Unmanned aerial vehicles (UAVs) are used to perform inspections, inspections, and monitoring tasks, replacing or supplementing traditional manual labor. This technology integrates advances in multiple fields, including drone platforms, automatic control technology, sensor technology, image processing technology, and data analysis and processing. Using drones for image acquisition reduces manual effort and increases the scope of image acquisition. The prior art closest to the present application is capable of using drones for image acquisition, but it still has certain limitations and deficiencies when applied in complex forest environments.

[0003] In the existing technology, first, there is a lack of specific optimization for forest environments. The existing technology mainly targets scenes such as substations and mountains, and does not fully consider the particularities of the forest environment, such as complex terrain, dense vegetation and other factors; second, the route planning lacks adaptability. Most methods rely on pre-set fixed routes and cannot be dynamically adjusted according to actual flight conditions, making it difficult to cope with uncertainties in the forest environment. Third, the image acquisition process lacks intelligent control. Existing methods usually use fixed parameters for shooting, and fail to automatically adjust the shooting parameters according to the characteristics of different areas, affecting image quality. Fourth, the data processing efficiency is low. Some methods need to transmit a large amount of data to the ground station for processing, which increases the communication burden and reduces real-time performance. Fifth, safety considerations are insufficient. In complex forest environments, existing methods lack effective obstacle avoidance and emergency processing mechanisms, posing safety risks. Summary of the Invention

[0004] The main purpose of the present invention is to provide a method and system for autonomous trajectory planning and image acquisition of unmanned aerial vehicles (UAVs), so as to solve the technical problem that the current UAV technical solutions can only simply control the UAV to fly along a fixed route and collect images according to a fixed pattern, and cannot meet the requirements of adaptive flight and image acquisition in complex terrain and complex environments.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is: a method for autonomous trajectory planning and image acquisition of a UAV, comprising the following steps:

[0006] S1: Obtain preset forest boundary data and generate an initial gridded flight area based on the forest boundary data; Based on the initial gridded flight area, combine terrain data and vegetation density data to obtain the current terrain undulation and vegetation distribution, and generate a three-dimensional gridded flight area;

[0007] S2: In the three-dimensional gridded flight area, an improved ant colony algorithm is used to plan an initial route that covers all grids and has the shortest flight distance. Current safety factors, including meteorological data and forest fire risk level data, are obtained, and the initial route is dynamically optimized based on the safety factors to generate a real-time route.

[0008] S3: The drone flies according to the real-time route and uses the onboard lidar to detect the surrounding environment in real time and perform obstacle avoidance.

[0009] S4: Adaptively adjust the camera's focal length, shutter speed, and ISO sensitivity to the optimal state based on the current drone's flight altitude, speed, and vegetation characteristics in the collection area, and control the camera to collect images;

[0010] S5: The collected images are input into the deep learning model deployed on the drone, which performs real-time image analysis and processing to identify areas that meet the preset feature conditions;

[0011] S6: Based on the recognition results, a list of areas requiring special attention is generated. The drone automatically plans additional routes based on the list of areas, collects and analyzes further images of the areas requiring special attention, and accurately identifies and locates areas that meet the preset characteristic conditions.

[0012] Step S1 specifically includes:

[0013] S101: Resampling the terrain data and vegetation density data based on the same resolution as the initial gridded flight area, including: assigning an elevation value to each grid point in the initial gridded flight area based on the terrain data to generate a preliminary three-dimensional grid; and adaptively adjusting the density of the preliminary three-dimensional grid based on the vegetation density data to increase the grid density in areas where the vegetation density is higher than a preset threshold.

[0014] S102: Calculate the safe flight altitude for each grid point; the safe flight altitude is the sum of the larger value of the terrain altitude, the vegetation altitude, and the minimum safe altitude;

[0015] S103: Perform three-dimensional Laplace smoothing on the grid, and update the height of each grid point by calculating the average height of the neighboring points around each grid point and combining it with the smoothing factor;

[0016] S104: Perform collision detection to ensure that the safe flight altitude of all grid points is higher than the surrounding terrain and vegetation, and adjust the altitude of grid points with potential collision risks;

[0017] S105: generating an optimized three-dimensional gridded flight area based on the grid data after adaptive density adjustment, safe flight altitude calculation, three-dimensional Laplace smoothing, and collision detection adjustment;

[0018] In step S2, an improved ant colony algorithm is used to plan an initial route that covers all grids and has the shortest flight distance, which specifically includes the following steps:

[0019] S201: constructing a three-dimensional pheromone matrix, the three-dimensional pheromone matrix including three-dimensional coordinate information of grid points; initializing parameters of the improved ant colony algorithm, including weights of a pheromone importance factor, a heuristic factor, and a height difference factor; and calculating an adaptive state transition probability, the adaptive state transition probability taking into account pheromone concentration, heuristic information, and a height difference factor;

[0020] Performing a path search based on the adaptive state transition probability; applying a local search strategy based on a 2-opt algorithm to the searched path, and optimizing the current path by exchanging two edges in the path;

[0021] S202: updating the three-dimensional pheromone matrix using a dual pheromone update mechanism, wherein the dual pheromone update mechanism combines global optimal information and local elite solution information;

[0022] S203: Dynamically adjust the parameters of the improved ant colony algorithm according to the current search results, including the weights of the pheromone importance factor, the heuristic factor, and the height difference factor; based on the updated parameters and the three-dimensional pheromone matrix, repeatedly execute the path search, local search strategy, pheromone update, and parameter adjustment steps until the preset termination condition is met; select the path with the shortest flight distance and covering all grid points from the multiple paths that meet the termination condition as the initial route.

[0023] In the preferred solution, the adaptive state transition probability formula is:

[0024] ;

[0025] Where, is the probability that ant k moves from the current grid point i to the next grid point j, Pheromone importance factor, is the weight of the heuristic factor, is the weight of the height difference factor, is the pheromone concentration from grid point i to grid point j, is the heuristic information, representing the expected degree from point i to point j; is the height difference factor, used to balance the change of flight altitude. are the coordinates of all possible next grid points.

[0026] In the preferred solution, in step S2, current safety factors are obtained, including meteorological data and forest fire risk level data, and the initial route is dynamically optimized based on the safety factors to generate a real-time route, specifically including:

[0027] S21: Real-time collection of meteorological parameters, including temperature, humidity, wind speed, and wind direction, and calculation of forest fire risk level data based on the forest fire meteorological index. Kriging spatial interpolation algorithm is used to process discrete meteorological data and forest fire risk level data to generate continuous meteorological parameter and fire risk level distribution maps.

[0028] S22: Construct a multi-objective optimization function including a flight safety scoring function, a mission efficiency function, and an energy consumption function, and solve the function to obtain an optimization result; wherein the safety scoring function considers fire risk level and wind speed factors, the mission efficiency function evaluates route coverage and flight time, and the energy consumption function is related to flight distance and altitude change; and use the fuzzy analytic hierarchy process to determine the weight coefficient of each objective in the multi-objective optimization function; and use an improved particle swarm optimization algorithm to solve the multi-objective optimization problem, wherein the improved particle swarm optimization algorithm introduces adaptive inertia weight and chaotic perturbation mechanism;

[0029] S23: Based on the optimization results, the initial route is dynamically adjusted to generate a real-time route that takes safety factors into consideration.

[0030] In the preferred embodiment, step S3 specifically includes:

[0031] The drone flies according to the real-time route and uses the lidar system to detect the surrounding environment in real time;

[0032] Use real-time obstacle avoidance algorithms to handle dynamic obstacles encountered during flight and obtain real-time obstacle avoidance results;

[0033] The flight path is adjusted accordingly based on real-time obstacle avoidance results.

[0034] In the preferred embodiment, step S5 specifically includes the following steps:

[0035] S501: Input the collected images into the deep learning model deployed on the drone;

[0036] S502: The deep learning model based on the Transformer architecture is a multi-task model that analyzes and processes the input image; the multi-task model includes an image reconstruction module, a Transformer encoder, a multi-scale feature fusion module and a task-specific head; the deep learning model is optimized using MixUp and CutMix data enhancement techniques, Dropout and L2 regularization, a cosine annealing learning rate scheduling strategy and label smoothing technology; the deep learning model is trained using a multi-task loss function, which includes classification loss, detection loss, segmentation loss and image reconstruction loss; and the deep learning model is used to identify areas that meet preset feature conditions.

[0037] In the preferred embodiment, step S6 specifically includes the following steps:

[0038] S601: Based on the recognition results of the deep learning model, an adaptive key area generation algorithm is used to generate a list of areas requiring special attention. The adaptive key area generation algorithm uses a weighted scoring function to calculate area scores. The weighted scoring function comprehensively considers recognition confidence, area size, fire risk level, and geographical location importance, and introduces a time decay factor to dynamically adjust the key areas of attention.

[0039] S602: Encode the generated list of key areas of interest using a compact binary format, wherein the binary format includes header information and multiple area data, each area data including center coordinates, a bounding box, a score, and attribute flags; and send the encoded list of key areas of interest to the drone.

[0040] S603: After receiving the list of key areas of concern, the UAV automatically plans a supplementary route using a real-time route planning method based on the improved A algorithm;

[0041] The improved A algorithm adopts a heuristic function that comprehensively considers distance, remaining battery power, communication quality, and obstacle avoidance; implements an online replanning mechanism and sets replanning trigger conditions based on changes in battery power, communication quality, and distance to obstacles; and the drone further collects images of areas requiring special attention along the planned supplementary route to achieve accurate identification and positioning of areas that meet the characteristics of the preset areas.

[0042] An autonomous UAV trajectory planning and image acquisition system, used to implement the autonomous UAV trajectory planning and image acquisition method, comprising:

[0043] The 3D grid generation module is used to obtain preset forest boundary data and generate an initial gridded flight area based on the forest boundary data. Based on the initial gridded flight area, the module combines terrain data and vegetation density data to obtain the current terrain relief and vegetation distribution, and generates a 3D gridded flight area.

[0044] The route planning module is used to plan an initial route that covers all grids and has the shortest flight distance in a three-dimensional gridded flight area using an improved ant colony algorithm. It also obtains current safety factors, including meteorological data and forest fire risk level data, and dynamically optimizes the initial route based on these safety factors to generate a real-time route.

[0045] The obstacle avoidance module uses the drone to fly according to the real-time route and uses the airborne laser radar to detect the surrounding environment in real time to perform obstacle avoidance.

[0046] The image acquisition module is used to adaptively adjust the camera's focal length, shutter speed, and ISO sensitivity to the optimal state based on the current UAV's flight altitude, speed, and vegetation characteristics of the acquisition area, and control the camera to acquire images;

[0047] The image processing module is used to input the collected images into the deep learning model deployed on the drone, perform real-time image analysis and processing, and identify areas that meet preset feature conditions;

[0048] The recognition and feedback module is used to generate a list of areas that need to be focused on based on the recognition results. The drone automatically plans supplementary routes based on the area list, conducts further image collection and analysis on the areas that need to be focused on, and realizes accurate recognition and positioning of areas that meet the preset feature conditions.

[0049] The present invention provides a method and system for autonomous trajectory planning and image acquisition of unmanned aerial vehicles. First, through intelligent three-dimensional grid division and dynamic route planning, the coverage and efficiency of feature area monitoring are significantly improved. Second, the integration of real-time meteorological data and historical feature data enhances the pertinence and adaptability of acquisition, which helps to obtain high-quality images in complex and changing environments. The combination of real-time processing and ground-based deep learning analysis greatly improves the accuracy and efficiency of feature extraction. Based on the adaptive supplementary sampling mechanism of key focus areas, rapid response and precise positioning of suspected feature areas are achieved, providing reliable data support for subsequent detailed analysis and decision-making. The present invention is not only suitable for various recognition tasks of preset features, but also has good scalability and flexibility, and can be adjusted and optimized according to different application requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0051] Figure 1 It is a flow chart of the method of the present invention;

[0052] Figure 2 This is a diagram of the deep learning model structure of the present invention;

[0053] Figure 3 It is a simulation diagram of the operation results of the method of the present invention. DETAILED DESCRIPTION

[0054] Example 1

[0055] like Figure 1-3 As shown, a method for autonomous trajectory planning and image acquisition of a UAV includes the following steps:

[0056] S1: Obtain preset forest boundary data and generate an initial gridded flight area based on the forest boundary data; based on the initial gridded flight area, combine terrain data and vegetation density data, and consider terrain undulation and vegetation distribution to generate a three-dimensional gridded flight area.

[0057] S2: In the three-dimensional gridded flight area, an improved ant colony algorithm is used to plan an initial route that covers all grids and has the shortest flight distance. The current safety factors, including meteorological data and forest fire risk level data, are obtained, and the initial route is dynamically optimized based on the safety factors to generate a real-time route.

[0058] S3: The drone flies according to the real-time route and uses the onboard lidar to detect the surrounding environment in real time and perform obstacle avoidance.

[0059] S4: According to the current UAV flight altitude, speed and vegetation characteristics of the collection area, the camera's focal length, shutter speed and ISO sensitivity are adaptively adjusted to the optimal state, and the camera device is controlled to collect images.

[0060] S5: The collected images are input into the deep learning model deployed on the drone, and the images are analyzed and processed in real time to identify areas that meet the preset feature conditions.

[0061] S6: Based on the recognition results, a list of areas that need to be focused on is generated. The drone automatically plans a supplementary route based on the area list, and further images are collected and analyzed for the areas that need to be focused on, so as to achieve accurate recognition and positioning of areas that meet the preset feature conditions.

[0062] In this embodiment, a three-dimensional gridded flight area is first generated, and an improved ant colony algorithm is used for route planning and dynamic optimization, real-time environmental detection and obstacle avoidance, and adaptive image acquisition is performed. The acquired images are input into a deep learning model deployed on the drone, and the images are analyzed in real time to identify areas that meet preset feature conditions. Further analysis is performed on key areas to ensure that the drone can fully cover the designated area, adapt to complex environments, and quickly respond to environmental changes and safety factors, thereby improving safety and flexibility. Combined with deep learning technology, the recognition rate of intelligent target areas is improved, and the efficiency and accuracy of data acquisition are improved.

[0063] The beneficial effects of the technical solution of the present invention are mainly reflected in the following aspects:

[0064] Intelligent 3D meshing and dynamic route planning improve monitoring coverage and efficiency. Combining onboard real-time processing with ground-based deep learning analysis enhances feature extraction accuracy. An adaptive supplementary sampling mechanism based on key areas enables rapid response and precise positioning of suspected feature areas.

[0065] In this embodiment, the preferred data specifically adopted are as follows:

[0066] In step S1, GPS data of the forest boundary is first obtained to generate an initial 100m × 100m grid. This is combined with DEM data (elevation 120–450m, slope 0–35 degrees) and vegetation density data (canopy density 0.4–0.8), and trilinear interpolation is used to generate a three-dimensional gridded flight area with a vertical interval of 50m.

[0067] In step S3, the drone flies along its real-time route at a cruising speed of 15 m / s. Its onboard LiDAR system (with a detection range of 100 m and a point cloud density of 50 points per square meter) scans its surroundings in real time. When an obstacle is detected, the obstacle avoidance algorithm generates an avoidance path within 0.1 seconds, ensuring a safe distance of at least 50 m from the obstacle.

[0068] After steps S1-S6, the system generates identification results and generates a list of key areas of concern, including three suspected pest and disease sites and two fire hazard sites. The drone automatically descends to a flight altitude of 30 meters and captures images of these areas again at a speed of 5 meters per second, with a collection interval of 2 seconds, to obtain more accurate monitoring data.

[0069] In the preferred embodiment, step S1 specifically includes:

[0070] S101: Resampling the terrain data and vegetation density data based on the same resolution as the initial gridded flight area, including: assigning an elevation value to each grid point in the initial gridded flight area based on the terrain data to generate a preliminary three-dimensional grid; and performing adaptive density adjustment on the preliminary three-dimensional grid based on the vegetation density data, increasing the grid density in areas where the vegetation density is higher than a preset threshold.

[0071] S102: Calculate the safe flight altitude for each grid point; the safe flight altitude is the sum of the larger value of the terrain altitude, the vegetation altitude, and the minimum safe altitude.

[0072] S103: Perform three-dimensional Laplace smoothing on the grid, and update the height of each grid point by calculating the average height of neighboring points around each grid point and combining it with a smoothing factor.

[0073] S104: Perform collision detection to ensure that the safe flight altitude of all grid points is higher than the surrounding terrain and vegetation, and adjust the altitude of grid points with potential collision risks.

[0074] S105: Generate an optimized three-dimensional gridded flight area based on the grid data after adaptive density adjustment, safe flight altitude calculation, three-dimensional Laplace smoothing processing and collision detection adjustment.

[0075] Furthermore, the system uses a rasterization algorithm to convert the forest boundary into an initial gridded flight area (50 m x 50 m). Terrain data (DEM) and vegetation density data (NDVI) are resampled to match the initial grid resolution. A three-dimensional grid is generated by assigning elevation values ​​to each grid point based on the terrain data. In areas where vegetation density exceeds 60%, the grid is subdivided into smaller 25 m x 25 m cells for adaptive density adjustment.

[0076] Calculate the safe flight altitude for each grid point as the greater of the terrain height plus the vegetation height or the minimum safe altitude (20 meters). Optimize the grid using 3D Laplace smoothing. Calculate the average altitude of each grid point's neighbors and apply a smoothing factor (0.5) to update the grid point's altitude. Repeat this process until the desired smoothing effect is achieved.

[0077] Collision detection is performed within a 100-meter radius to ensure that the safe flight altitude of each grid point is above the surrounding terrain and vegetation. For grid points with a collision risk, the altitude is adjusted to be above the highest surrounding point and a 5-meter safety margin is added. This process generates the final optimized 3D grid flight area.

[0078] In steps S101-S105, by introducing high-precision terrain data and vegetation density data, the generated three-dimensional gridded flight area accurately reflects the actual terrain characteristics and vegetation distribution, providing a reliable foundation for subsequent route planning. Secondly, the adaptive density adjustment mechanism ensures higher sampling accuracy in areas with dense vegetation, improving the targetedness and efficiency of data collection. Thirdly, the calculation of the safe flight altitude and collision detection steps significantly enhance flight safety and effectively reduce the risk of drone operation. Finally, three-dimensional Laplace smoothing optimizes the smoothness of the flight path, which helps improve flight stability and data collection quality.

[0079] In the preferred solution, an improved ant colony algorithm is used to plan an initial route that covers all grids and has the shortest flight distance, which specifically includes the following steps:

[0080] S201: Construct a three-dimensional pheromone matrix: The three-dimensional pheromone matrix contains the three-dimensional coordinate information of the grid points; initialize the parameters of the improved ant colony algorithm, including the weights of the pheromone importance factor, the heuristic factor and the height difference factor; calculate the adaptive state transition probability, and the adaptive state transition probability takes into account the pheromone concentration, the heuristic information and the height difference factor.

[0081] Path search is performed based on the adaptive state transition probability; a local search strategy based on the 2-opt algorithm is applied to the searched path, and the current path is optimized by exchanging two edges in the path.

[0082] S202: Update the three-dimensional pheromone matrix using a dual pheromone update mechanism: The dual pheromone update mechanism combines global optimal information and local elite solution information.

[0083] S203: Dynamically adjust the parameters of the improved ant colony algorithm according to the current search results, including the weights of the pheromone importance factor, the heuristic factor, and the height difference factor; based on the updated parameters and the three-dimensional pheromone matrix, repeatedly execute the path search, local search strategy, pheromone update, and parameter adjustment steps until the preset termination condition is met; select the path with the shortest flight distance and covering all grid points from the multiple paths that meet the termination condition as the initial route.

[0084] Furthermore, after the three-dimensional gridded flight area is generated, the system uses an improved ant colony algorithm to plan the initial route.

[0085] First, the system constructs a three-dimensional pheromone matrix, which contains the three-dimensional coordinate information and initial pheromone concentration of each grid point.

[0086] Next, the system initializes the parameters of the improved ant colony algorithm, including the weights of the pheromone importance factor α (e.g., 1.0), the heuristic factor β (e.g., 2.0), and the height difference factor γ (e.g., 0.5).

[0087] During path search, the system calculates an adaptive state transition probability. This probability takes into account three factors: the pheromone concentration from the current grid point to the next possible grid point, the Euclidean distance between the two points (as a heuristic), and the height difference. The system calculates the transition probability using a weighted sum of these three factors, with weights α, β, and γ, respectively. For example, if the pheromone concentration from the current grid point to a neighboring point is 2.0, the Euclidean distance is 50 meters, and the height difference is 10 meters, a composite score can be calculated; a higher score indicates a greater transition probability.

[0088] Based on the calculated transition probabilities, the system conducts a path search. Each ant begins at a randomly selected starting point and selects the next grid point based on the transition probability until all grid points are covered or the maximum number of steps is reached. During this process, the system records each ant's path and total flight distance.

[0089] The system optimizes each ant's path using a local search strategy based on the 2-opt algorithm. It randomly selects two non-adjacent edges in the path and swaps them. If the new path is shorter, the change is accepted. This process is repeated until no better local improvement can be found.

[0090] The system uses a dual pheromone update mechanism to update the three-dimensional pheromone matrix, combining the global optimal and local elite solution information. The pheromone concentration is increased by 0.1 for the global optimal path and 0.05 for the local optimal solution.

[0091] When no optimal solution is found in successive iterations, the system dynamically adjusts the α and β values ​​to balance exploration and exploitation. Path search, local search strategy, pheromone update, and parameter adjustment are repeated until the maximum number of iterations (1000) is reached or there is no significant improvement after 50 consecutive iterations.

[0092] Finally, from the paths that meet the termination conditions, the path with the shortest flight distance and covering all grid points is selected as the initial route. The algorithm improves the convergence speed and solution quality through the three-dimensional pheromone matrix, adaptive state transition probability and dual pheromone update mechanism.

[0093] In this embodiment, step S2 is the route planning stage.

[0094] In a three-dimensional grid space, an improved ant colony algorithm was used with a population size of 50, an iteration count of 200, and a pheromone volatility coefficient of 0.3. The algorithm considered constraints such as flight altitude and route spacing to generate an initial route with a total length of approximately 380 kilometers. Real-time weather station data (temperature 26°C, relative humidity 55%, easterly wind 4 m / s, moderate fire risk level) was acquired and combined with Kriging spatial interpolation to generate a continuous meteorological field. The route was dynamically optimized to generate a real-time route that avoided high-risk areas. The specific data and results are shown in Table 1 below.

[0095] Table 1 Route data and optimization results

[0096]

[0097] In the preferred solution, the adaptive state transition probability formula is:

[0098] ;

[0099] Where, is the probability that ant k moves from the current grid point i to the next grid point j, Pheromone importance factor, is the weight of the heuristic factor, is the weight of the height difference factor, is the pheromone concentration from grid point i to grid point j, is the heuristic information, representing the expected degree from point i to point j; is the height difference factor, used to balance the change of flight altitude. are the coordinates of all possible next grid points.

[0100] In this embodiment, the algorithm adaptively selects the optimal path based on historical information (pheromones), current state (distance), and environmental characteristics (altitude change). By dynamically adjusting the values ​​of α, β, and γ, the algorithm balances global exploration and local exploitation during the search process, thereby improving the efficiency of finding the optimal solution. In the preferred solution, current safety factors, including meteorological data and forest fire risk level data, are obtained. The initial route is dynamically optimized based on these safety factors to generate a real-time route, specifically including:

[0101] S21: Real-time collection of meteorological parameters, including temperature, humidity, wind speed and direction, and calculation of forest fire risk level data based on the forest fire meteorological index; use the Kriging spatial interpolation algorithm to process discrete meteorological data and forest fire risk level data to generate continuous meteorological parameter and fire risk level distribution maps.

[0102] S22: Construct a multi-objective optimization function including a flight safety scoring function, a mission efficiency function, and an energy consumption function, and solve it to obtain the optimization result; among them, the safety scoring function takes into account the fire risk level and wind speed factors, the mission efficiency function evaluates the route coverage and flight time, and the energy consumption function is related to the flight distance and altitude change; and use the fuzzy hierarchical analysis method to determine the weight coefficient of each objective in the multi-objective optimization function; adopt an improved particle swarm optimization algorithm to solve the multi-objective optimization problem, and the improved particle swarm optimization algorithm introduces adaptive inertia weight and chaotic perturbation mechanism.

[0103] S23: Based on the optimization results, the initial route is dynamically adjusted to generate a real-time route that takes safety factors into consideration.

[0104] In the preferred embodiment, step S3 specifically includes:

[0105] The drone flies according to the real-time route and uses the lidar system to detect the surrounding environment in real time.

[0106] Use real-time obstacle avoidance algorithm to handle dynamic obstacles encountered during flight and obtain real-time obstacle avoidance results.

[0107] The flight path is adjusted accordingly based on real-time obstacle avoidance results.

[0108] Furthermore, through the above steps S2-S3, the system collects key meteorological parameters including temperature, relative humidity, wind speed and wind direction in real time through the automatic weather station network deployed in the forest area.

[0109] Based on the collected meteorological data, the system uses the forest fire meteorological index system to calculate the forest fire risk level data.

[0110] The system uses the Kriging spatial interpolation algorithm to process these discrete data and generate continuous meteorological parameters and fire risk level distribution maps.

[0111] Next, the system constructs a multi-objective optimization function that considers flight safety, mission efficiency, and energy consumption. The safety score function considers fire risk and wind speed. The mission efficiency function evaluates route coverage and flight time. The energy consumption function accounts for flight distance and altitude changes.

[0112] The system uses fuzzy analytic hierarchy process to determine the weight coefficient of each objective in the multi-objective optimization function.

[0113] To solve multi-objective optimization problems, the system uses an improved particle swarm optimization algorithm. This algorithm incorporates adaptive inertia weights. It also incorporates a chaotic perturbation mechanism, which applies small random perturbations to some particles during each iteration, enhancing the algorithm's ability to escape local optima.

[0114] Based on the optimization results, the system dynamically adjusts the initial route to generate a real-time route that takes safety factors into consideration.

[0115] The system controls the drone to fly according to the real-time route, and uses the high-precision lidar system on board to detect the surrounding environment in real time.

[0116] During flight, the system uses real-time obstacle avoidance algorithms to deal with dynamic obstacles that may be encountered.

[0117] Based on the real-time obstacle avoidance results, the system makes local adjustments to the flight path to ensure the safe flight of the drone.

[0118] In steps S2-S3 of this embodiment, first, by introducing real-time meteorological data and forest fire risk level data, the safety and adaptability of UAV route planning are significantly improved, enabling the system to quickly respond to environmental changes. Secondly, the design of the multi-objective optimization function achieves a balance between safety, task efficiency, and energy consumption, ensuring the comprehensive performance of the route. Thirdly, the improved particle swarm optimization algorithm improves the solution efficiency and result quality, enabling the system to complete route optimization in a short time. Finally, the introduction of the real-time obstacle avoidance system greatly enhances the autonomous flight capability of the UAV and improves flight safety. This dynamic optimization and real-time control method is not only suitable for forest monitoring, but can also be extended to other UAV application scenarios in complex environments, such as disaster monitoring, urban inspections, etc., and has broad practical value and promotion prospects.

[0119] Steps S4-S5 adaptively adjust the camera to the optimal state based on the current UAV's flight altitude, speed, and vegetation characteristics of the collection area, control the camera to capture images, input the images into the deep learning model, perform real-time image analysis and processing, and identify areas that meet the preset characteristic conditions;

[0120] In this embodiment, during the flight of the drone, based on the current flight parameters and vegetation characteristics, the camera parameter control algorithm adjusts the focal length to 35mm, the shutter speed to 1 / 1000 second, and the ISO sensitivity to 400. 4K resolution is used, and an image is captured every 5 seconds to ensure image clarity and an overlap rate of at least 85%, as shown in Table 2 below.

[0121] Table 2 Image quality control index table

[0122]

[0123] like Figure 3 As shown in the figure, the adaptive camera parameters (focal length, shutter speed, and ISO) demonstrate the system's ability to adaptively adjust camera settings based on fire risk and environmental conditions in different areas. This dynamic adjustment enables the drone in this embodiment to automatically adjust its imaging equipment during monitoring, resulting in clearer and more accurate data collected under varying environmental conditions.

[0124] In the preferred embodiment, step S5 specifically includes the following steps:

[0125] S501: Input the collected images into the deep learning model deployed on the drone.

[0126] S502: The deep learning model based on the Transformer architecture is a multi-task model that analyzes and processes the input image; the multi-task model includes an image reconstruction module, a Transformer encoder, a multi-scale feature fusion module, and a task-specific head; the deep learning model is optimized using MixUp and CutMix data augmentation techniques, Dropout and L2 regularization, a cosine annealing learning rate scheduling strategy, and label smoothing technology; the deep learning model is trained using a multi-task loss function, which includes classification loss, detection loss, segmentation loss, and image reconstruction loss; and the deep learning model is used to identify areas that meet preset feature conditions.

[0127] In this example, the deep learning model first feeds the captured image into a preprocessing module for normalization and data augmentation. MixUp blends the two images with weights of 0.7 and 0.3, while CutMix replaces the 224×224 image by cropping a 128×128 region.

[0128] The image reconstruction module adopts an encoder-decoder structure, contains 5 convolution blocks, uses a dropout rate of 30% and an L2 regularization coefficient of 0.0001.

[0129] The Transformer encoder uses a multi-head self-attention mechanism with 8 attention heads and a hidden layer dimension of 512. The position encoding uses sine-cosine encoding, and the encoder contains 6 layers of standard Transformer blocks.

[0130] The multi-scale feature fusion module uses a feature pyramid structure to fuse features of different scales obtained from the Transformer encoder. The top-down path uses 1×1 convolutions for channel adjustment, and the lateral connection uses 3×3 convolutions for feature fusion. This ultimately generates feature maps containing multi-scale semantic information, ranging from original size to 1 / 16 size.

[0131] The task-specific head consists of three branches: classification, detection, and segmentation. The classification branch uses global average pooling followed by a fully connected layer to output class probabilities.

[0132] The training process uses a cosine annealing learning rate strategy, with an initial learning rate of 0.001 and a periodic adjustment every 100 epochs. The label smoothing coefficient is set to 0.1. The weights of each task loss in the multi-task loss function are adaptively adjusted based on the performance on the validation set. The initial weight ratio is classification: detection: segmentation: reconstruction = 1:1:1:0.5. The specific image preprocessing results are shown in Table 3 below.

[0133] Table 3 Image preprocessing specifications and effects

[0134]

[0135] The beneficial effects are: the use of multi-task learning improves the model's generalization ability; the Transformer structure enhances the feature extraction effect; multi-scale feature fusion improves the detection accuracy of targets of different scales; data enhancement and adaptive loss weight optimization ensure the stability of model training and achieve efficient forest monitoring.

[0136] In the preferred embodiment, step S6 specifically includes the following steps:

[0137] S601: Based on the recognition results of the deep learning model, an adaptive key area generation algorithm is used to generate a list of areas that need to be focused on. The adaptive key area generation algorithm uses a weighted scoring function to calculate the regional score. The weighted scoring function comprehensively considers the recognition confidence, regional area, fire risk level and geographical location importance, and introduces a time decay factor to dynamically adjust the key focus areas.

[0138] S602: Encode the generated list of key areas of interest using a compact binary format, where the binary format includes header information and multiple area data, each area data including center coordinates, bounding box, score, and attribute flags; and send the encoded list of key areas of interest to the drone.

[0139] S603: After receiving the list of key areas of concern, the drone automatically plans a supplementary route using a real-time route planning method based on the improved A algorithm.

[0140] The improved A algorithm adopts a heuristic function that comprehensively considers distance, remaining battery power, communication quality, and obstacle avoidance; implements an online replanning mechanism and sets replanning trigger conditions based on changes in battery power, communication quality, and distance to obstacles; and the drone conducts further image acquisition of areas that require special attention along the planned supplementary route to achieve accurate identification and positioning of areas that meet the characteristics of the preset areas.

[0141] like Figure 3 As shown in the figure, it includes path planning and fire risk avoidance, error tolerance control, precise monitoring and secondary data collection in high-priority areas, and adaptive adjustment of cameras. The forest area simulated using test data includes data such as terrain height, vegetation density, and fire risk level. The information in the figure specifically includes:

[0142] Figure 3The middle green dashed line represents the ideal path, running from the lower left corner to the upper right corner of the map, marking the ideal flight path for the drone when there is no fire risk. The solid red line represents the actual optimized path used in this example. Deviations from the actual path are calculated to ensure the drone effectively avoids high-risk fire areas without affecting the overall smoothness of the route. The optimized path verifies the effectiveness of the route optimization by avoiding high-risk fire areas. High-priority target points are also marked within the fire risk areas, further verifying the drone's ability to identify and accurately locate key areas.

[0143] The colored background includes:

[0144] 1) Altitude (first color bar): This uses terrain height data to simulate terrain variations. The color changes represent different altitudes, from low to high, reflecting the rolling terrain that a route may encounter.

[0145] 2) Vegetation density (second color bar): Green transparent areas indicate different vegetation cover densities, which increase the complexity of the flight path.

[0146] 3) Fire Danger Level (third color bar): Displays the distribution of fire danger levels, ranging from green (low fire danger) to red (high fire danger), so that the drone path can avoid fire risks.

[0147] 4) Error Band (orange-filled area): This represents the acceptable error range for the route, based on a 5-meter error range above and below the ideal path. It shows the tolerance for deviation between the ideal and optimized paths, verifying that the drone's navigation system accurately avoids high-risk areas within the preset error tolerance.

[0148] Further includes:

[0149] Secondary collection path (purple dashed line): The secondary collection path around the high-priority area is used for subsequent precise monitoring, which shows that drones need to pay special attention to key areas and conduct secondary collection to ensure the accuracy of fire risk monitoring data.

[0150] Camera parameters (focal length, shutter speed, ISO): Near the high-priority areas, each area displays the camera's adaptive acquisition parameters. The differentiated settings of the parameters demonstrate the system's ability to adaptively adjust acquisition under different environmental conditions, enabling the drone to collect high-quality images and data based on environmental changes.

[0151] High-priority areas (red markers): Several areas with high fire risk levels were selected along the ideal path, marked as pink dots, and avoided by the actual route. These areas were set as target areas for key fire risk detection.

[0152] In summary, it can be concluded that the optimized path of this embodiment effectively avoids high-priority areas and achieves the precise avoidance requirements for fire risk areas; the addition of secondary collection paths shows that this embodiment automatically detects and performs secondary collection tasks in required areas based on the fire risk level, further improving the accuracy of fire risk monitoring and greatly improving the accurate recognition rate and positioning accuracy that meet the characteristics of the preset areas.

[0153] In this embodiment, an adaptive key area generation algorithm is used to generate a list of areas requiring special attention based on the recognition results of the ground station's deep learning model. This algorithm uses a weighted scoring function that comprehensively considers multiple factors. The recognition confidence reflects the reliability of the model's judgment of the area, typically ranging from 0 to 1. The area considers the scale of the problem; the geographical importance considers the strategic value of the area. The algorithm also introduces a time decay factor to dynamically adjust the key areas of attention.

[0154] After generating the list of key areas of interest, it is transmitted to the drone in a compact binary format. This format consists of a header and multiple region data. The header may contain information such as a version number, a timestamp, and the total number of regions. Each region data contains the following information: center coordinates, bounding box, score, and attribute flags.

[0155] After receiving the list of key areas of interest, the drone automatically plans a supplementary route for further image acquisition. This approach uses a real-time route planning method based on the improved A algorithm. The improved A algorithm uses a heuristic function to evaluate path costs, taking into account multiple factors: distance, remaining battery power, communication quality, and obstacle avoidance.

[0156] The route planning algorithm also implements an online replanning mechanism, setting replanning trigger conditions based on changes in battery level, communication quality, and distance to obstacles.

[0157] Following the planned supplementary route, the drone will collect further images of areas requiring special attention. During this process, the drone may adjust its flight altitude and speed to obtain clearer images.

[0158] In step S6 of this embodiment, the adaptive key area generation algorithm improves monitoring efficiency and accuracy; the compact binary encoding reduces the amount of data transmission; the route planning based on the improved A algorithm ensures flight safety and monitoring effectiveness; and the online replanning mechanism enhances system adaptability.

[0159] Example 2

[0160] Further illustrate with reference to Example 1, Figure 2The structure shown is a UAV autonomous trajectory planning and image acquisition system, which is used to implement the UAV autonomous trajectory planning and image acquisition method in Example 1, including:

[0161] The three-dimensional grid generation module is used to obtain preset forest boundary data and generate an initial gridded flight area based on the forest boundary data; based on the initial gridded flight area, the current terrain undulation and vegetation distribution are obtained by combining terrain data and vegetation density data to generate a three-dimensional gridded flight area.

[0162] The route planning module is used to plan an initial route that covers all grids and has the shortest flight distance in a three-dimensional gridded flight area using an improved ant colony algorithm; obtain current safety factors, which include meteorological data and forest fire risk level data, and dynamically optimize the initial route based on the safety factors to generate a real-time route.

[0163] The obstacle avoidance module uses the drone to fly according to the real-time route and adopts the airborne laser radar to detect the surrounding environment in real time to perform obstacle avoidance.

[0164] The image acquisition module is used to adaptively adjust the camera's focal length, shutter speed, and ISO sensitivity to the optimal state according to the current UAV's flight altitude, speed, and vegetation characteristics of the acquisition area, and control the camera device to acquire images.

[0165] The image processing module is used to input the collected images into the deep learning model deployed on the drone, perform real-time image analysis and processing, and identify areas that meet preset feature conditions.

[0166] The recognition and feedback module is used to generate a list of areas that need to be focused on based on the recognition results. The drone automatically plans supplementary routes based on the area list, conducts further image collection and analysis on the areas that need to be focused on, and realizes accurate recognition and positioning of areas that meet the preset feature conditions.

[0167] The above embodiments are merely preferred technical solutions of the present invention and should not be construed as limiting the present invention. The scope of protection of the present invention shall be the technical solutions set forth in the claims, including equivalent alternatives to the technical features of the technical solutions set forth in the claims. In other words, equivalent alternatives and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for autonomous trajectory planning and image acquisition of a UAV, characterized in that: The following steps are involved: S1: Obtain preset forest boundary data, generate an initial gridded flight area based on the forest boundary data, and based on the initial gridded flight area, obtain the current terrain undulation and vegetation distribution by combining terrain data and vegetation density data, and generate a three-dimensional gridded flight area; S2: In the three-dimensional gridded flight area, an improved ant colony algorithm is used to plan an initial route that covers all grids and has the shortest flight distance. Current safety factors, including meteorological data and forest fire risk level data, are obtained, and the initial route is dynamically optimized based on the safety factors to generate a real-time route. S3: The drone flies according to the real-time route and uses the onboard lidar to detect the surrounding environment in real time and perform obstacle avoidance. S4: Adaptively adjust the camera's focal length, shutter speed, and ISO sensitivity to the optimal state based on the current drone's flight altitude, speed, and vegetation characteristics in the collection area, and control the camera to collect images; S5: The collected images are input into the deep learning model deployed on the drone, which performs real-time image analysis and processing to identify areas that meet the preset feature conditions; S6: Based on the recognition results, a list of areas requiring special attention is generated. The drone automatically plans additional routes based on the list of areas, collects and analyzes further images of the areas requiring special attention, and accurately identifies and locates areas that meet the preset characteristic conditions. Step S1 specifically includes: S101: Resampling the terrain data and vegetation density data based on the same resolution as the initial gridded flight area, including: assigning an elevation value to each grid point in the initial gridded flight area based on the terrain data to generate a preliminary three-dimensional grid; and adaptively adjusting the density of the preliminary three-dimensional grid based on the vegetation density data to increase the grid density in areas where the vegetation density is higher than a preset threshold. S102: Calculate the safe flight altitude for each grid point; the safe flight altitude is the sum of the larger value of the terrain altitude, the vegetation altitude, and the minimum safe altitude; S103: Perform three-dimensional Laplace smoothing on the grid, and update the height of each grid point by calculating the average height of the neighboring points around each grid point and combining it with the smoothing factor; S104: Perform collision detection to ensure that the safe flight altitude of all grid points is higher than the surrounding terrain and vegetation, and adjust the altitude of grid points with potential collision risks; S105: generating an optimized three-dimensional gridded flight area based on the grid data after adaptive density adjustment, safe flight altitude calculation, three-dimensional Laplace smoothing, and collision detection adjustment; In step S2, an improved ant colony algorithm is used to plan an initial route that covers all grids and has the shortest flight distance, which specifically includes the following steps: S201: constructing a three-dimensional pheromone matrix, the three-dimensional pheromone matrix including three-dimensional coordinate information of grid points; initializing parameters of the improved ant colony algorithm, including weights of a pheromone importance factor, a heuristic factor, and a height difference factor; and calculating an adaptive state transition probability, the adaptive state transition probability taking into account pheromone concentration, heuristic information, and a height difference factor; Performing a path search based on the adaptive state transition probability; applying a local search strategy based on a 2-opt algorithm to the searched path, and optimizing the current path by exchanging two edges in the path; S202: updating the three-dimensional pheromone matrix using a dual pheromone update mechanism, wherein the dual pheromone update mechanism combines global optimal information and local elite solution information; S203: Dynamically adjust the parameters of the improved ant colony algorithm according to the current search results, including the weights of the pheromone importance factor, the heuristic factor, and the height difference factor; based on the updated parameters and the three-dimensional pheromone matrix, repeatedly execute the path search, local search strategy, pheromone update, and parameter adjustment steps until the preset termination condition is met; select the path with the shortest flight distance and covering all grid points from the multiple paths that meet the termination condition as the initial route.

2. The autonomous trajectory planning and image acquisition method for unmanned aerial vehicles according to claim 1, characterized in that: The formula for adaptive state transition probability is: ； Where, is the probability that ant k moves from the current grid point i to the next grid point j, Pheromone importance factor, is the weight of the heuristic factor, is the weight of the height difference factor, is the pheromone concentration from grid point i to grid point j, is the heuristic information, representing the expected degree from point i to point j; is the height difference factor, used to balance the change of flight altitude. are the coordinates of all possible next grid points.

3. The autonomous trajectory planning and image acquisition method for unmanned aerial vehicles according to claim 1, characterized in that: In step S2, current safety factors are obtained, including meteorological data and forest fire risk level data, and the initial route is dynamically optimized based on the safety factors to generate a real-time route, specifically including: S21: Real-time collection of meteorological parameters, including temperature, humidity, wind speed, and wind direction, and calculation of forest fire risk level data based on the forest fire meteorological index. Kriging spatial interpolation algorithm is used to process discrete meteorological data and forest fire risk level data to generate continuous meteorological parameter and fire risk level distribution maps. S22: Construct a multi-objective optimization function including a flight safety scoring function, a mission efficiency function, and an energy consumption function, and solve the function to obtain an optimization result; wherein the safety scoring function considers fire risk level and wind speed factors, the mission efficiency function evaluates route coverage and flight time, and the energy consumption function is related to flight distance and altitude change; and use the fuzzy analytic hierarchy process to determine the weight coefficient of each objective in the multi-objective optimization function; and use an improved particle swarm optimization algorithm to solve the multi-objective optimization problem, wherein the improved particle swarm optimization algorithm introduces adaptive inertia weight and chaotic perturbation mechanism; S23: Based on the optimization results, the initial route is dynamically adjusted to generate a real-time route that takes safety factors into consideration.

4. The autonomous trajectory planning and image acquisition method for unmanned aerial vehicles according to claim 3, characterized in that: Step S3 specifically includes: The drone flies according to the real-time route and uses the lidar system to detect the surrounding environment in real time; Use real-time obstacle avoidance algorithms to handle dynamic obstacles encountered during flight and obtain real-time obstacle avoidance results; The flight path is adjusted accordingly based on real-time obstacle avoidance results.

5. The autonomous trajectory planning and image acquisition method for unmanned aerial vehicles according to claim 1, characterized in that: Step S5 specifically includes the following steps: S501: Input the collected images into the deep learning model deployed on the drone; S502: The deep learning model based on the Transformer architecture is a multi-task model that analyzes and processes the input image; the multi-task model includes an image reconstruction module, a Transformer encoder, a multi-scale feature fusion module and a task-specific head; the deep learning model is optimized using MixUp and CutMix data enhancement techniques, Dropout and L2 regularization, a cosine annealing learning rate scheduling strategy and label smoothing technology; the deep learning model is trained using a multi-task loss function, which includes classification loss, detection loss, segmentation loss and image reconstruction loss; and the deep learning model is used to identify areas that meet preset feature conditions.

6. The autonomous trajectory planning and image acquisition method for unmanned aerial vehicles according to claim 1, characterized in that: Step S6 specifically includes the following steps: S601: Based on the recognition results of the deep learning model, an adaptive key area generation algorithm is used to generate a list of areas requiring special attention. The adaptive key area generation algorithm uses a weighted scoring function to calculate area scores. The weighted scoring function comprehensively considers recognition confidence, area size, fire risk level, and geographical location importance, and introduces a time decay factor to dynamically adjust the key areas of attention. S602: Encode the generated list of key areas of interest using a compact binary format, wherein the binary format includes header information and multiple area data, each area data including center coordinates, a bounding box, a score, and attribute flags; and send the encoded list of key areas of interest to the drone. S603: After receiving the list of key areas of concern, the UAV automatically plans a supplementary route using a real-time route planning method based on the improved A algorithm; The improved A algorithm adopts a heuristic function that comprehensively considers distance, remaining battery power, communication quality, and obstacle avoidance; implements an online replanning mechanism and sets replanning trigger conditions based on changes in battery power, communication quality, and distance to obstacles; and the drone further collects images of areas requiring special attention along the planned supplementary route to achieve accurate identification and positioning of areas that meet the characteristics of the preset areas.

7. An autonomous trajectory planning and image acquisition system for unmanned aerial vehicles, characterized in that: The method for autonomous trajectory planning and image acquisition of a UAV according to any one of claims 1 to 6 comprises: The 3D grid generation module is used to obtain preset forest boundary data and generate an initial gridded flight area based on the forest boundary data. Based on the initial gridded flight area, the module combines terrain data and vegetation density data to obtain the current terrain relief and vegetation distribution, and generates a 3D gridded flight area. The route planning module is used to plan an initial route that covers all grids and has the shortest flight distance in a three-dimensional gridded flight area using an improved ant colony algorithm. It also obtains current safety factors, including meteorological data and forest fire risk level data, and dynamically optimizes the initial route based on these safety factors to generate a real-time route. The obstacle avoidance module uses the drone to fly according to the real-time route and uses the airborne laser radar to detect the surrounding environment in real time to perform obstacle avoidance. The image acquisition module is used to adaptively adjust the camera's focal length, shutter speed, and ISO sensitivity to the optimal state based on the current UAV's flight altitude, speed, and vegetation characteristics of the acquisition area, and control the camera to acquire images; The image processing module is used to input the collected images into the deep learning model deployed on the drone, perform real-time image analysis and processing, and identify areas that meet preset feature conditions; The recognition and feedback module is used to generate a list of areas that need to be focused on based on the recognition results. The drone automatically plans supplementary routes based on the area list, conducts further image collection and analysis on the areas that need to be focused on, and realizes accurate recognition and positioning of areas that meet the preset feature conditions.

Images