A single unmanned aerial vehicle autonomous path planning method in a network-free environment

Abstract

The application discloses a single unmanned aerial vehicle autonomous path planning method in a network-free environment, which comprises the following steps: 1) according to the distance between the unmanned aerial vehicle and surrounding obstacles, obstacles on a flight road are divided to obtain the state level of the surrounding environment of the unmanned aerial vehicle; 2) a targeted flight method is formulated according to the state level of the surrounding environment of the unmanned aerial vehicle; 3) multi-target optimization is introduced into the path planning method, and the flight method of the unmanned aerial vehicle is comprehensively controlled through three different factors, i.e., average flight speed, flight decision number and flight energy consumption; and 4) according to the multi-target optimization, a reward function in a DDPG method is further optimized. The method can effectively cope with different flight environments, and the path planning capability of the unmanned aerial vehicle is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) path planning technology, and more particularly to a method for autonomous path planning of a single UAV in a network-free environment. Background Technology

[0002] Commercial drones primarily rely on GPS and base station communication. In environments where network access is unavailable, their autonomous path planning capabilities are quite weak. Therefore, most drones can only operate in low-risk, line-of-sight environments. In previous research, vision-based methods have been crucial for real-time path planning in drones due to the low cost and rich sensory information capabilities of camera sensors. Visual sensing has wide applications in autonomous systems, mainly in target tracking and rapid obstacle avoidance. However, drones relying solely on visual sensors are limited by extreme weather conditions such as fog and rain. Besides cameras, lidar sensors are also a vital tool. They can create 3D topographic maps of landscapes and provide detailed terrain descriptions through captured point clouds, offering advantages such as stability, low cost, and ease of deployment. In earlier stages, Simultaneous Localization and Mapping (SLAM) was a mature autonomous path navigation method, utilizing lidar sensors in GPS-restricted or congested indoor environments. However, due to the uneven distribution of point clouds, re-meshing and interpolation processes are often required to generate a digital surface model (DSM). This rasterization-induced information loss significantly impacts the accuracy of the DSM. Furthermore, SLAM methods require map construction after acquiring data using LiDAR sensors. In complex, dynamic environments, map modeling is a challenging task. While it can achieve the goal of reaching a designated target point, it remains difficult to quickly complete the task in completely unknown environments. Of course, combining LiDAR and cameras can leverage their respective strengths to accomplish the task. UAVs can fly in unknown environments where GPS is limited, enabling detection and search missions.

[0003] In typical continuous control domains, actions are real values ​​with a large action space, making it difficult for agents to learn directly from pixel-level inputs to perform complex operational tasks. Reinforcement learning (RL) is frequently used due to its powerful learning, prediction, and decision-making capabilities. Besides applying RL to UAV flight decisions to learn from previous successful experiences, it can effectively find better decisions. UAVs have demonstrated improved learning capabilities thanks to the emergence of deep reinforcement learning (DRL). DRL combines RL with deep learning (DL), enabling UAVs to achieve increasingly better results in complex environments, tasks, and repetitive flights. In several simulated continuous control scenarios, a novel DRL method called Deep Deterministic Gradient Method (DDPG) has shown good performance. DDPG is used to train UAVs to reach spatially dispersed moving or static objects in a 3D urban environment, with collision penalties and customized rewards used to shorten the distance between the UAV and the target point.

[0004] Autonomous planning methods for drones have been studied for decades, even before the concept of artificial intelligence was introduced. Combinations of various sensors and basic methods can effectively perform path planning in general environments, even without GPS signals. For example, well-performing SLAM methods can achieve autonomous exploration of the path forward. However, it cannot work in completely unknown environments. This is because the method requires pre-imported map information and the comparison and planning of path information.

[0005] With the development of artificial intelligence, Regression Learning (RL) and Deep Learning (DL) are widely used to achieve autonomous flight decisions for unmanned aerial vehicles (UAVs). Methods such as Q-learning, DQN, and DDPG have enabled autonomous path planning for UAVs in simple simulation environments. However, some problems remain, such as slow convergence speed and a lack of consideration for real-world environments. These issues are reflected in the selection and simplification of obstacles and the analysis of flight path conditions. The simplification of obstacle shapes in the penalty function does not match real-world states, and the collision depth in the penalty function cannot be obtained in the real world. Furthermore, obstacles are simplified to basic rectangles, and the graphics are two-dimensional, failing to accurately depict the complexity and dynamics of the real world; simultaneously, some researchers have not considered the few obstacle areas in the real world where UAVs can fly at high speeds. Therefore, they cannot improve overall operational and learning efficiency.

[0006] In summary, we have studied an intelligent method for UAV autonomous path planning that considers MOO (Multi-Objective Optimization) in real-world network unavailable environments. Summary of the Invention

[0007] In view of this, the present invention discloses a method for autonomous path planning of a single UAV in a network-free environment, so as to enable the UAV to cope with different flight environments in an unknown network-free environment and improve the path planning capability of the UAV.

[0008] The technical solution provided by this invention is specifically a method for autonomous path planning of a single unmanned aerial vehicle (UAV) in a network-free environment, comprising:

[0009] 1) Based on the distance between the UAV and surrounding obstacles, the obstacles on the flight path are divided to obtain the state level of the UAV's surrounding environment;

[0010] 2) Develop targeted flight methods based on the environmental conditions surrounding the UAV;

[0011] 3) Introduce multi-objective optimization into the path planning method, and use three different factors—average flight speed, number of flight decisions, and flight energy consumption—to comprehensively control the flight of the UAV.

[0012] 4) Based on multi-objective optimization, further optimize the reward function in the DDPG method.

[0013] Furthermore, in step 1), information about the drone's surroundings is obtained through radar sensors to determine the state of the drone's environment.

[0014] Further, step 2) specifically involves: setting a distance threshold of 10 meters between the drone and obstacles, where a safe flight zone is defined as the drone not detecting any obstacles within a circle with a radius of 10 meters; the flight method within this safe zone is as follows: using v max Fly in a straight line at high speed, then turn off the camera;

[0015] When the distance between the drone and the obstacle is between 5 and 10 meters, it is considered a warning zone. The flight method within this warning zone involves the use of both radar and camera sensors. The drone needs to determine the area where the target point is located and assign weights to the velocity components of that area. The drone's velocity is set to 0.75*v. max Depth images are acquired using a depth-sensing camera.

[0016] When the distance between the drone and an obstacle is less than 5 meters, the drone is considered to be in a danger zone. The flight method in the danger zone is as follows: the optimized DDPG method is used for path planning; the drone is trained to avoid obstacles by acquiring depth images; and the drone uses a forced mechanism to prevent collisions through a lidar sensor.

[0017] Furthermore, in step 3)

[0018] The drone detects its distance to obstacles in real time and adopts different flight methods based on the detected distance; the average flight speed can be expressed as:

[0019]

[0020] Where D k (l s , l e ) is the coordinates of the UAV. s and endpoint coordinates l e The distance between them, K is the total number of target points in the flight, t k This is the total flight time;

[0021] The total number of flight decisions can be expressed as:

[0022]

[0023] Where, N k It is the number of flight decisions between every two target points;

[0024] Total energy consumption can be expressed as:

[0025]

[0026] Where E(t) k It is the flight energy consumption between every two target points;

[0027] Therefore, the overall multi-objective optimization can be summarized as follows:

[0028]

[0029] stv(t)∈[0, v max ],t∈[0,T] (b)

[0030] θ(t)∈[-π, π], N col (t)≤N max (c)

[0031] Where N max It is the maximum number of times the device collides with an obstacle.

[0032] Furthermore, the reward function in the optimized DDPG method is specifically as follows:

[0033]

[0034] Where r ue r su r con This represents three optimization objectives: maximizing average flight speed, minimizing decision-making, and minimizing flight energy consumption; specifically...

[0035] Described as:

[0036] r ue =exp{(-||D(l u (t), l e )-D(l u (t-1), l e (6)

[0037] r su =-exp{-N col (t)}-1 (7)

[0038]

[0039] Where the norm ||D(l) u (t), l e )-D(l u (t-1), l e || represents the absolute value of the internal data, D(l) u (t), l e () is the coordinate l of the UAV at time t. u (t) and endpoint coordinates l e The distance between them, N col E(t) represents the number of collisions that occurred with the drone, and E(t) represents the energy consumed by the drone during flight.

[0040] The remaining r sta Represented as:

[0041]

[0042] If the drone reaches the target point and completes the mission, then r sta The value is 0.1, and 0 in all other cases.

[0043] This invention provides a single UAV autonomous path planning method in a network-free environment. First, the method takes into account the complexity of the real environment and divides different flight conditions by the distance between the UAV and obstacles, effectively coping with different flight environments and greatly improving the path planning capability of the UAV.

[0044] Secondly, this invention adaptively schedules sensors in different scenarios, shutting down unnecessary camera sensors in areas with fewer obstacles to reduce energy consumption. Simultaneously, when using the optimized DDPG method, radar sensors are invoked to force flight using physical means, preventing drone collisions. To further improve flight energy efficiency, both intelligent and non-intelligent methods are adaptively scheduled to reduce computation time and computational load.

[0045] Thirdly, to improve the autonomous path planning capability of UAVs under different environments and requirements, this invention employs a flight control method that comprehensively manages UAV flight based on three different factors: average flight speed, number of flight decisions, and flight energy consumption. This achieves switching between three different requirements—speed, stability, and low energy consumption—which is of great significance for UAV flight in various situations.

[0046] Finally, to consider drone flight control methods in real-world environments, where flight decision-making requires speed and computational complexity, this invention optimizes the traditional DDPG method by replacing random experience selection with selective experience filtering, resulting in faster convergence. Furthermore, to more closely approximate data acquired in real-world environments, the reward function is further optimized.

[0047] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit the disclosure of the present invention. Attached Figure Description

[0048] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This invention discloses an application scenario for the autonomous path planning method for unmanned aerial vehicles (UAVs) in a network-free environment.

[0051] Figure 2 A schematic diagram showing the results of both the method of the prior art and the method of the present invention successfully completing the task, as provided in the embodiments of the present invention.

[0052] Figure 3 This is a schematic diagram of the flight trajectory of a drone after implementing various methods, provided in the embodiments of the present invention.

[0053] Figure 4 A schematic diagram illustrating the impact of changing the experience pool parameters on the evaluation index in the method provided by the embodiments of the present invention.

[0054] Figure 5 This diagram illustrates the impact of changing the batch parameter on the evaluation metrics in the method provided by the embodiments of the present invention.

[0055] Figure 6A schematic diagram illustrating the impact of changing the actor and critic learning rate parameters on the evaluation metrics in the method provided by the embodiments of the present invention.

[0056] Figure 7 A flowchart of an autonomous path planning method for unmanned aerial vehicles (UAVs) in a network-free environment, as disclosed in an embodiment of the present invention. Detailed Implementation

[0057] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of systems consistent with some aspects of the invention as detailed in the appended claims.

[0058] In real-world environments where networks are unavailable, this implementation plan provides an autonomous UAV path planning method to improve the overall operation and learning efficiency of UAVs. This method uses optimized artificial intelligence to adjust different sensors according to different scenarios. Furthermore, to adapt to different environments, a multi-objective optimization problem is introduced into the method to achieve optimal performance.

[0059] This implementation plan specifically includes:

[0060] A method for autonomous path planning of a single unmanned aerial vehicle (UAV) in a network-free environment includes:

[0061] 1) Based on the distance between the UAV and surrounding obstacles, the obstacles on the flight path are divided to obtain the state level of the UAV's surrounding environment;

[0062] 2) Develop targeted flight methods based on the environmental conditions surrounding the UAV;

[0063] 3) Introduce multi-objective optimization into the path planning method, and use three different factors—average flight speed, number of flight decisions, and flight energy consumption—to comprehensively control the flight of the UAV;

[0064] 4) Based on multi-objective optimization, further optimize the reward function in the DDPG method.

[0065] In step 1), information about the drone's surroundings is obtained through radar sensors to determine the state of the drone's environment.

[0066] Step 1 considers the complexity of flight paths (in real-world environments, flight paths are complex and unpredictable, with obstacle distribution constantly changing). This implementation scheme divides the flight path into sections, employing different flight methods for different sections. The division is based on the distance between the UAV and surrounding obstacles. Information about the UAV's surroundings is obtained through radar sensors, revealing the state of the UAV's environment. The maximum radar detection range of this invention is 10 meters. If the UAV detects no obstacles within a 10-meter radius circle, it represents a safe flight zone. When the distance is less than 5 meters, the UAV is in a danger zone and requires careful handling. When the distance is between 5 and 10 meters, it is a warning zone, and the UAV needs to reduce its flight speed to avoid collisions.

[0067] When the distance between the drone and an obstacle is less than 5 meters, the drone is considered to be in a danger zone. The flight method in the danger zone is as follows: the optimized DDPG method is used for path planning; the drone is trained to avoid obstacles by acquiring depth images; and the drone uses a forced mechanism to prevent collisions through lidar sensors.

[0068] The selection of multi-sensor scheduling and flight control methods in this implementation is crucial. Different sensors need to be activated for the UAV in different scenarios. Furthermore, in simple environments, a simple flight control method is sufficient; employing intelligent methods would increase the UAV's computational load and energy consumption.

[0069] The radar sensor data obtained in step 2) is analyzed to develop a targeted flight strategy. The distance threshold between the UAV and obstacles is set to 10 meters. When the UAV is in a safe area, it can fly rapidly, i.e., in a straight line at a speed of 20 m / s. During this period, the camera is turned off to reduce wasted computing resources. In unsafe sections, the cooperation of both radar and camera sensors is considered to reduce computation and improve flight efficiency. When the distance between the UAV and obstacles is between 5 and 10 meters, the UAV needs to determine the area where the target point is located and add a certain weight to the velocity component of the corresponding area. Simultaneously, it is necessary to reduce the UAV's speed, setting it to...

[0070] At a speed of 0.75*20m / s, the depth camera is activated at any time to acquire depth images. When the distance between the drone and obstacles is less than 5 meters, the drone can use an optimized DDPG method for path planning. By acquiring depth images, the drone can effectively train its obstacle avoidance capabilities. Simultaneously, the lidar sensor is also used as a mandatory mechanism to prevent drone collisions.

[0071] To enable the drone's flight method to adapt to different scenario requirements, this implementation plan incorporates multi-objective optimization. Simply adjusting the weights of different requirements allows it to adapt to varying environments. In step 3), the drone detects its distance to obstacles in real time and adopts different flight methods based on the detected distance.

[0072] This step, through different weight ratios, allows for switching between multiple environments and requirements, effectively improving the adaptability of the invention. The objective of this invention in this project is to maximize average flight speed while minimizing the number of decisions and flight energy consumption. To achieve real-time path planning, the UAV must detect its distance to obstacles in real time and adopt different flight methods based on the detected distance. Furthermore, when planning the UAV's flight path, the number of flight decisions and energy consumption must be considered. The average flight speed can be expressed as:

[0073]

[0074] Where D k (l s ,l e ) is the coordinates of the UAV. s and endpoint coordinates l e The distance between them, K is the total number of target points in the flight, t k That is the total flight time.

[0075] The total number of flight decisions can be expressed as:

[0076]

[0077] Where, N k It represents the number of flight decisions between every two target points.

[0078] Total energy consumption can be expressed as:

[0079]

[0080] Where E(t) k It is the flight energy consumption between every two target points.

[0081] Therefore, the overall multi-objective optimization problem can be reduced to:

[0082]

[0083] stv(t)∈[0, v max ],t∈[0,T] (b)

[0084] θ(t)∈[-π, π], N col (t)≤N max (c)

[0085] Where N max It is the maximum number of times the device collides with an obstacle.

[0086] Existing intelligent methods have failed to be widely applied in real-world scenarios, largely due to their slow convergence speed. This invention optimizes the original DDPG method. The reward function plays a crucial role in intelligent methods; therefore, this implementation optimizes the reward function to adapt to complex environments and improve convergence speed. The reward function is expressed as follows:

[0087]

[0088] Where r ue r su r con This represents three optimization objectives: maximizing average flight speed, minimizing decision costs, and minimizing flight energy consumption. Their specific descriptions are as follows:

[0089] r ue =exp{(-||D(l u (t), l e )-D(l u (t-1), l e (6)

[0090] r su =-exp{-N col (t)}-1 (7)

[0091]

[0092] Where the norm ||D(l) u (t), l e )-D(l u (t-1), l e || represents the absolute value of the internal data, D(l) u (t), l e () is the coordinate l of the UAV at time t. u (t) and endpoint coordinates l e The distance between them, N col E(t) represents the number of collisions that occurred with the drone, and E(t) represents the energy consumed by the drone during flight.

[0093] The remaining r sta Represented as:

[0094]

[0095] The drone takes off from its starting position and remains in flight until it reaches its destination. Faster flight and fewer decisions are rewarded, while higher flight energy consumption is penalized. The appropriate weight for each link is ω. ue ω su ω con It also considers the global state reward r sta If the drone reaches the target point and completes the mission, the value is set to 0.1; otherwise, it is 0. The experience gained from each flight cycle has a different value for the entire experience pool, so these values ​​must be differentiated. This can be achieved by using r... sta To achieve this, when the drone reaches its final state, its experience should remain organized, and r should be added to the reward for each step. sta These experiences are then added to the experience pool. Therefore, each successful experience plays a more significant role in the experience pool, improving learning efficiency.

[0096] Unlike traditional DDPGs, this invention does not arbitrarily select experience tuples when optimizing the learning network. To improve learning efficiency, a method of selecting appropriate experiences is used to learn more effective experience tuples. For example, when avoiding obstacles, experiences closer to the obstacle are more valuable. Therefore, past experiences can be selected to optimize flight performance.

[0097] The proposed method for autonomous path planning of unmanned aerial vehicles (UAVs) in real-world environments without network connectivity was simulated in PyCharm. The UAV's flight trajectory was obtained using the AirSim platform. The method was compared with the original DDPG and a DDPG with a representative experience replay mechanism. Experimental results show that the proposed method outperforms the original global DDPG by 50% and converges 30% faster than the global experience replay mechanism DDPG. Furthermore, to better verify the performance of different methods, the original DDPG with a multi-sensor adaptive approach and the experience replay mechanism DDPG were compared. A simulated environment was created by combining Unreal Engine 4 (UE4) with AirSim and Visual Studio 2019 to generate a realistic environment and manage the UAV's flight. Free-form cave materials in UE4 were used to approximate the real scene. Obstacles in the cave are statically scattered, irregularly shaped, unevenly sized, and fixedly distributed. The UAV's starting and target points are located within a square with sides of 750 meters. On one hand, when the number of collisions reaches a threshold, the UAV returns to its starting position and flies again. On the other hand, after successfully reaching the destination, the UAV returns to its starting position and enters the next cycle. Its maximum speed is 20 meters per second, and its flight altitude is limited to 1 to 5 meters. There are four target points, and the drone's flight path is arc-shaped. Furthermore, these schemes are evaluated using three metrics: average flight speed, number of flight decisions, and flight energy consumption.

[0098] Appendix Figure 2 The results of successfully completing the task using various methods are shown. The UAV was controlled to take one step every 0.25 seconds, with the target position fixed. The UAV's flight speed was determined by a normal distribution with parameters μ = 15 and σ = 1. (See attached diagram.) Figure 2 As shown in (a), the average flight speed of the UAV is plotted. It can be seen that the three methods equipped with lidar sensors have a faster average flight speed throughout the entire flight. This is because the UAV can fly at a higher speed on straight segments, thus increasing the overall flight speed. Among the three methods equipped with lidar sensors, the present invention reaches a new equilibrium more quickly, reflecting its convergence. The reason for this effect is that the scheme can filter learning experiences and learn from past experiences more effectively. Appendix Figure 2 (b) shows the number of successful flight decisions made using these methods, reflecting the correctness and effectiveness of UAV decision-making. (See Appendix) Figure 2 (b) It can be seen that the method without a LiDAR sensor requires more decisions. This is because the UAV relies entirely on the global DDPG method. The UAV needs to go through numerous trials and errors, and cannot effectively understand the road conditions ahead, thus increasing the number of decisions. In contrast, UAVs equipped with LiDAR sensors can effectively reduce the number of decisions. Appendix Figure 2 (c) shows the flight energy consumption of five different methods. It can be seen that the flight energy consumption of the UAV equipped with a lidar sensor is significantly reduced. Due to the reduction in the number of decision-making steps and the UAV's flight time, energy consumption can inevitably be minimized.

[0099] Appendix Figure 3 The flight trajectories of the drones after implementing various methods are depicted. Each sub-graph shows a pitch angle of 35° at the top and -90° at the bottom. Overall, [the following text appears to be unrelated and possibly machine-generated]: Appendix Figure 3 (a) and appendix Figure 3 (c) The overall flight trajectory of the drone without a lidar sensor fluctuates more, especially to the left of marker 1. This is because there are almost no obstacles in this area, and drones equipped with lidar sensors can use them to determine the distance between the drone and obstacles. If the drone is in a safe state, its flight speed will increase, the number of decision-making steps will decrease, and therefore the trajectory fluctuation will be smaller. The drone signal trajectories at markers 1 and 2 in each figure can reflect the drone's obstacle avoidance capabilities. From the appendix... Figure 3 (a) and appendix Figure 3 As can be seen from marker 1 in (c), the drone is far from the obstacle. Furthermore, the flight path is zigzag and the transitions are unnatural. (See attached...) Figure 3 (b) and appendix Figure 3In (d), the drone is equipped with a lidar sensor, resulting in a closer distance between the drone and the obstacle at marker point 1, and a more natural flight path transition. For the drone employing the solution of this invention, the trajectory transition at marker point 1 is smoother. Simultaneously, the distance between the drone and the obstacle is very close, effectively shortening the flight distance. This performance is due to the force exerted by the lidar sensor. Because there is a non-line-of-sight connection between the starting and ending points, the drone's trajectory should approximate an arc. Furthermore, the shortest distance is a line segment between two points, so the drone needs to fly close to the obstacle to shorten the flight distance.

[0100] The size of the experience pool and sampling batch, as well as the learning rates of the actor and critic networks, are key factors determining the performance of the DDPG method. To demonstrate the convergence of the method in this invention, the values ​​of three evaluation criteria are given after changing the three core parameters. By fixing two of the three parameters and setting the third parameter to a range of variation, their impact on performance can be observed more comprehensively. The experience pool is divided into three sizes: 8000, 10000, and 12000, and the sampling batch sizes are 32, 64, and 128, respectively. The basic learning rates for the actor and critic networks are α_actor = 0.00001 and α_critic = 0.00005, respectively, with the other two groups multiplied or divided by a factor of 100. (Appendix) Figure 4 , 5 Tables 6 and 7 respectively show the impact of changing the three parameters—experience pool, batch, actor, and critic—on the evaluation metrics. Figure 4 It can be seen that when the number of cycles is small, the average flight speed of the smaller experience pool is higher, and the number of decisions and flight energy consumption are lower. As the number of cycles increases, the values ​​of the three different experience pool sizes tend to be consistent. Figure 5 In the early stages, both larger and smaller batches performed poorly. This is because smaller batches store less successful experience, while larger batches represent a smaller proportion of successful experience. Regarding the learning rate, although a larger learning rate allows for faster learning and reaching a new equilibrium, it also results in greater volatility. Figure 6 This is further highlighted in (c). As can be seen from the figure, when the UAV successfully completes a limited number of flight missions, the differences in results are significant. However, at a certain later stage, the scheme in this invention can effectively converge and eventually reach a stable value.

[0101] This invention provides a solution for autonomous path planning of unmanned aerial vehicles (UAVs) in real-world environments without network connectivity, addressing the problems of existing methods that fail to consider the real environment, multi-sensor scheduling and method selection, and slow convergence speed of intelligent methods. Experimental results show that the solution provided by this invention can improve convergence speed by 30% compared to other solutions.

[0102] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

Claims

1. A method for autonomous path planning of a single unmanned aerial vehicle (UAV) in a network-free environment, characterized in that, include: 1) Based on the distance between the UAV and surrounding obstacles, the obstacles on the flight path are divided to obtain the state level of the UAV's surrounding environment; 2) Develop targeted flight methods based on the environmental conditions surrounding the UAV; 3) Introduce multi-objective optimization into the path planning method, and control the flight of the UAV through a combination of multiple factors, including average flight speed, number of flight decisions and flight energy consumption; 4) Based on multi-objective optimization, further optimize the reward function in the DDPG method; In step 3), the drone detects the distance between itself and obstacles in real time and adopts different flight methods based on the detected distance; Average flight speed can be expressed as: in The coordinates of the drone and endpoint coordinates The distance between them, K is the total number of target points in the flight. This is the total flight time; The total number of flight decisions can be expressed as: in, It is the number of flight decisions between every two target points; Total energy consumption can be expressed as: in It is the flight energy consumption between every two target points; Therefore, the overall multi-objective optimization can be summarized as follows: in, That is the maximum flight speed of the drone. The number of drone collisions, T, represents the maximum flight time. It is the maximum number of collisions with obstacles; The reward function in the optimized DDPG method is specifically as follows: in This represents three optimization objectives: maximizing average flight speed, minimizing decision-making, and minimizing flight energy consumption; specifically described as follows: Where the norm Represents the absolute value of internal data. Represents the coordinates of the UAV at time t. and endpoint coordinates The distance between them Indicates the number of collisions involving the drone. This indicates the energy consumed during drone flight; The rest Represented as: If the drone reaches the target point and completes the mission, then The value is 0.1, and 0 in all other cases. Represents the global state reward.

2. The method for autonomous path planning of a single unmanned aerial vehicle in a network-free environment according to claim 1, characterized in that, In step 1), information about the drone's surroundings is obtained through radar sensors to determine the state of the drone's environment.

3. The method for autonomous path planning of a single unmanned aerial vehicle in a network-free environment according to claim 1, characterized in that, Step 2) Specifically: Set the distance threshold between the drone and the obstacle as follows: meters, where the drone is in a radius of If no obstacles are found within the circle of meters, it is considered a safe flight zone; The flight method in the safe flight area is as follows: Fly in a straight line at high speed, then turn off the camera; When the distance between the drone and the obstacle is arrive When the distance is between [distance] meters, it is a warning zone. The flight method within this warning zone is as follows: Using both radar and camera sensors in cooperation, the UAV needs to determine the area where the target point is located and add weights to the velocity components of that area. The UAV's speed is set to [speed value]. Depth images are acquired using a depth-sensing camera. When the distance between the drone and the obstacle is less than If the distance is 1 meter, it indicates that the drone is in a dangerous area. The flight method in the dangerous area is as follows: the optimized DDPG method is used for path planning; the drone is trained to avoid obstacles by acquiring depth images. Drones use lidar sensors as a forced mechanism to prevent collisions.

Images