An unmanned aerial vehicle optimal path planning method and system combined with a neural network

Abstract

The embodiment of the application discloses a kind of unmanned aerial vehicle optimal path planning method and system combined with neural network, it is related to unmanned aerial vehicle autonomous navigation and path planning technical field, can be planned to complex scene, and improve the efficiency and quality of path planning.The present application comprises: generating three-dimensional grid map data according to preset map parameters, the three-dimensional grid map data is as the sample of neural network module training;According to the heuristic function of the path planning module that the cost increment of neural network module output is updated;The flight path of unmanned aerial vehicle is generated by the updated path planning module.The present application can be used in low-altitude economic related application scenarios.

Description

Technical Field

[0001] This invention relates to the field of autonomous navigation and path planning technology for unmanned aerial vehicles (UAVs), specifically to an optimal path planning method and system for UAVs that incorporates neural networks. Background Technology

[0002] With the increasing application of drones in complex environments such as disaster relief, logistics delivery, and urban inspection, their path planning capabilities have become a key limiting factor. Traditional path planning algorithms, such as the A* algorithm, often suffer from suboptimal path planning, low computational efficiency, and poor adaptability in complex 3D environments due to overly simplistic heuristic functions. Especially in scenarios with complex terrain and varied obstacles, traditional methods struggle to quickly generate safe and efficient optimal paths. Therefore, how to plan for complex scenarios and improve the efficiency and quality of path planning has become a research topic that needs to be addressed. Summary of the Invention

[0003] The embodiments of the present invention provide a method and system for optimal path planning of unmanned aerial vehicles (UAVs) that combines neural networks, which can plan for complex scenarios and improve the efficiency and quality of path planning.

[0004] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0005] Firstly, a method for optimal path planning for unmanned aerial vehicles (UAVs) that incorporates neural networks includes:

[0006] S1. Generate three-dimensional raster map data according to preset map parameters, and use the three-dimensional raster map data as samples for training the neural network module;

[0007] S2. Update the heuristic function of the path planning module according to the cost increment output by the neural network module. The neural network module includes four input channels, which correspond to the obstacle and terrain cost, the target position, the current node position, and the next candidate node position, respectively. During the process of updating the heuristic function, the cost prediction function output by the neural network module corresponds to a 4-way parallel feature decoupling branch.

[0008] S3. Generate the drone's flight path using the updated path planning module.

[0009] Secondly, an optimal path planning system for unmanned aerial vehicles (UAVs) that incorporates neural networks includes: a map generation module, a neural network module, and a path planning module.

[0010] The map generation module is used to generate three-dimensional raster map data according to preset map parameters, and the three-dimensional raster map data is used as a sample for training the neural network module.

[0011] The neural network module is used to update the heuristic function of the path planning module based on the cost increment output by the neural network module. The neural network module includes four input channels, which correspond to the obstacle and terrain cost, the target position, the current node position, and the next candidate node position, respectively. During the process of updating the heuristic function, the cost prediction function output by the neural network module corresponds to a 4-way parallel feature decoupling branch.

[0012] The path planning module is used to generate the flight path of the UAV through the updated path planning module.

[0013] In this embodiment of the invention, the map generation module generates 3D raster maps containing various terrain features in batches; the neural network module adopts a lightweight CNN structure, with input including four-channel information: obstacle and terrain cost, target position, current node, and next candidate node, and output as the local movement cost increment; the path planning module introduces the cost increment predicted by CNN into the A* algorithm framework, dynamically optimizes the heuristic function, and generates a flight path that better fits the actual terrain. Compared with the traditional A* algorithm, which suffers from low planning efficiency and non-optimal paths in complex 3D environments due to overly simple heuristic functions, this embodiment improves the quality and efficiency of path planning, has good terrain adaptability and generalization performance, is suitable for complex scenarios such as disaster relief and logistics distribution, and can be extended to other mobile robots and autonomous driving fields. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 Technical roadmap of the algorithm provided in the embodiments of the present invention;

[0016] Figure 2 A schematic flowchart provided for embodiments of the present invention;

[0017] Figure 3 The neural network framework provided for embodiments of the present invention Figure 1 ;

[0018] Figure 4 The neural network framework provided for embodiments of the present invention Figure 2 ;

[0019] Figure 5 This is a schematic diagram illustrating the execution flow of the algorithm running in the map generation module provided in an embodiment of the present invention;

[0020] Figure 6 and Figure 7 The planning effect diagram provided for the embodiment of the present invention;

[0021] Figure 8 This is a schematic diagram of the method flow provided in an embodiment of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Embodiments of the present invention will be described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in the specification of the present invention means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0023] The purpose of this invention is to address the problems of poor adaptability, low planning efficiency, and suboptimal paths in existing UAV path planning algorithms under complex disaster environments. It provides an optimal UAV path planning algorithm that incorporates a Convolutional Neural Network (CNN). By introducing a CNN-based dynamic optimization heuristic function, the efficiency and quality of path planning are improved. Figure 1 , 2 As shown, this embodiment designs an optimal path planning scheme for unmanned aerial vehicles (UAVs) that combines neural networks. The scheme includes:

[0024] Build a benchmark testing environment, set up a 3D grid map, and collect information including obstacle distribution, terrain elevation, and target point location;

[0025] Design and train a CNN model with four-channel grid information as input and cost increment as output. The training data comes from the real cost of multiple path planning in a simulation environment.

[0026] The trained CNN model is embedded into the A* algorithm to replace the original heuristic function and achieve dynamic weight adjustment. The Manhattan distance and Euclidean distance in the heuristic function can be selected according to the actual scenario. The cost increment output by the CNN is used to correct the heuristic estimate to make it more in line with the actual terrain cost. The algorithm supports multi-target point path planning. By expanding the number of input layer channels, multiple target point information is introduced to achieve global optimal path search.

[0027] Path planning is performed in a test environment to evaluate path length, planning time, and success rate, and compared with traditional algorithms such as A*, particle swarm optimization, and ant colony optimization. Specifically, this can be implemented as follows: Figure 8 The method flow shown includes:

[0028] S1. Generate three-dimensional raster map data according to preset map parameters, and the three-dimensional raster map data is used as a sample for training the neural network module; wherein, according to the preset raster three-dimensional raster map dataset, the three-dimensional raster map data is generated according to the preset size, and the terrain features in the raster three-dimensional raster map dataset include at least urban buildings, ruins and bridges.

[0029] S2. Update the heuristic function of the path planning module based on the cost increment output by the neural network module; wherein, the neural network module is a lightweight convolutional neural network, including four input channels, which correspond to obstacle and terrain cost, target position, current node position, and next candidate node position, respectively. The CNN neural network structure designed in this embodiment has four input channels: obstacle and terrain cost layer, target position layer, current node layer, and next position node layer, and the output is the predicted cost increment; the mean squared error is used as the loss function, and the network parameters are optimized through backpropagation so that the cost prediction weights output by the CNN approximate the actual movement cost; by using a raster 3D raster map dataset constructed for multi-terrain environments, the mapping relationship between different terrain features and weights can be mastered by training the CNN model.

[0030] S3. Generate the drone's flight path using the updated path planning module.

[0031] Before S2, it also includes: a path planning module based on the A* algorithm that is associated with the cost increment, and the heuristic function of the path planning module is... The heuristic function for the initial path planning module is... Because the heuristic function of the A* algorithm is too simple and cannot fully adapt to various disaster scenarios—specifically, when there are too many 3D map grids, the A* algorithm struggles to find the optimal path—this invention adds a cost increment to the heuristic function. The new heuristic function takes the form: . It is the cost estimate from the initial state to the target state via state n. It is the actual cost of going from the initial state to state n in the state space. It is the estimated cost of the optimal path from state n to the target state. This represents the cost increment predicted by the convolutional neural network in the neural network module.

[0032] The states in the heuristic function correspond to the position coordinates of the UAV, where the starting position of the UAV is... The drone's current location is , and These are the x-coordinate and y-coordinate values ​​of the starting position, respectively. and These are the x and y coordinates of the current location, respectively. The Manhattan distance is: The Euclidean distance is: .

[0033] In each update, the neural network module outputs a cost prediction function. Corresponding to 4 parallel feature decoupling branches, such as Figure 3 As shown, the fully connected layer pointed to by the distance feature branch outputs the distance weight, the fully connected layer pointed to by the height feature branch outputs the height weight, the fully connected layer pointed to by the steering feature branch outputs the steering weight, and the fully connected layer pointed to by the obstacle feature branch outputs the obstacle weight.

[0034] Updated heuristics for: , , ,in, This indicates the current location and status of the drone. This indicates the next node position status of the drone. and Here are two modified cost functions. and For state n and state The corresponding flight costs, The coordinates of the next target point, , These are the x-axis and y-axis coordinates of the next target point, respectively.

[0035] Furthermore, , , ,in, Let x, y, and z be the coordinates of the node at the theoretical optimal distance, respectively. The coordinates of the nearest obstacle to the current position. , , These are the x-axis, y-axis, and z-axis coordinates of the nearest obstacle to the current position, respectively. For distance weighting function, This represents the maximum range of movement allowed in a single step. This is the z-axis coordinate of the current position. The z-axis coordinate of the next target point. For the shift weight function, Let the obstacle weight function be... , , and These are the predicted cost increments for the distance weight function, height weight function, turning weight function, and obstacle weight function, respectively.

[0036] In the neural network module, a loss function L is set to ensure the reliability of the generated prediction weights, and its functional form is as follows: ,in, The basic prediction loss is used to measure the error between the predicted cost increment and the actual value. To guide loss for safety, calculations are based on the distance between the current node and the obstacle; The smoothness regularization loss is calculated based on the angle of change in the direction of movement; The physical constraint loss is calculated based on the climb angle of the local path and the dynamic limit of the UAV. These are the weighting coefficients for each loss term. This embodiment optimizes the loss function by employing multiple constraints to further enhance its flexibility.

[0037] This embodiment also provides an optimal path planning system for unmanned aerial vehicles (UAVs) that incorporates neural networks, including: a map generation module, a neural network module, and a path planning module;

[0038] The map generation module is used to generate 3D raster map data according to preset map parameters. This 3D raster map data serves as a training sample for the neural network module. The map generation module batch-generates a large number of maps required for training the neural network. The module is designed to randomly (or according to settings) generate various obstacle environments within a 200m×200m×40m area and record the generated maps for subsequent neural network training. The raster 3D map dataset contains various terrain features such as city buildings, ruins, and bridges, used to train the network to adapt to complex environments. In practical applications, the map generation module can be modularly edited. It generates maps by searching a map library, which is independently constructed and does not rely on sensor data acquisition. Furthermore, it is primarily used for model training. For example, first, building feature modules, ruin feature modules, and bridge feature modules are defined and numbered. Then, obstacles in the map are determined using random numbers, and node coordinates are randomly selected as obstacle center points to construct a complete map.

[0039] The neural network module is used to update the heuristic function of the path planning module based on the cost increment output by the neural network module; wherein, the neural network module mainly includes an input layer and an output layer, the input layer includes four channels, and the output is the predicted cost increment;

[0040] The path planning module is used to generate the UAV's flight path through the updated path planning module. The path planning module includes two parts: a path planning algorithm and a path output interface. The path planning algorithm modifies the heuristic function of the A* algorithm by introducing the cost increment of convolutional neural network prediction into the heuristic function, forming a new heuristic function. The path output interface verifies the generated algorithm on various maps to obtain the final planning result.

[0041] This embodiment improves the feedback part of the algorithm in the neural network. Existing solutions mainly use the algorithm to correct the neural network, but do not feed the neural network back to the algorithm itself to optimize it. This limits the path planning effect of the neural network. This embodiment, however, uses an auxiliary algorithm mode to improve the planning effect of the algorithm itself. Specifically, as follows... Figure 1 As shown, the heuristic function is iterated through a neural network, and the activation function is set to the absolute value of the difference between the activation function and the total distance of the optimal path, thus obtaining a better heuristic function. The cost increment output by the CNN is also considered. The heuristic function is dynamically updated to ensure that the path fits the terrain cost and to generate the optimal static path.

[0042] Combination Figure 3 and Figure 4As shown, in this embodiment, the CNN model adopts a convolutional encoder structure. It uses a 4-channel input, corresponding to obstacle and terrain costs, target location, current node, and next candidate node, with an input size of 11×11 local grid (the perception range centered on the current node). The convolutional encoder uses a 3-layer convolutional structure to progressively extract environmental features. Each layer includes convolution, batch normalization, ReLU activation function (a non-linear correction unit used to introduce non-linearity, enabling it to learn complex input-output relationships), and pooling operations. The number of channels increases from 16 to 64, achieving multi-scale feature extraction. Finally, the cost prediction value is output through a fully connected layer. The Dropout function (a regularization technique whose core function is to randomly "shut down" some neurons in the neural network during training, thus preventing the model from over-relying on specific neuron or feature combinations, thereby significantly improving the model's generalization ability) is used to add noise, and the dropout rate p is set to 0.3 to prevent overfitting. During training, the Adam optimizer (whose function is to automatically adjust the learning speed of each parameter to improve learning efficiency) is used, with a learning rate α set to 0.001 and a channel size of 32.

[0043] Combination Figure 5 As shown, the map generation module in this embodiment first calibrates specific terrain and sets up a terrain library containing cities, ruins, mountains, bridges, etc. Then, after setting the map size, it randomly selects specific terrain from the terrain library and adds it. After generating the terrain, it is then stored in the map dataset.

[0044] Combination Figure 6 , Figure 7As shown, in this embodiment, the generated algorithm model is applied to some randomly generated scenarios to verify its planning effect. Compared with the traditional A* algorithm, the method used in this invention has a relatively shorter total path length and time. Comparing the two algorithms simultaneously on ten maps, the average time required for the traditional A* algorithm to complete planning on the same map is 35.75 seconds, while the average time required for the method used in this invention is 26.59 seconds. For each map, the planning time of the method used in this invention is relatively uniform (around 30 seconds), avoiding the situation in the traditional A* algorithm where some maps have very long planning times while others have very short planning times. Subtracting the path lengths planned using the traditional A* algorithm from the path lengths planned using the method of this invention on the ten maps used in this experiment yielded the following values: -8.03m, -8.60m, -9.47m, -7.72m, -3.60m, -0.07m, -7.47m, -4.53m, -0.41m, and -0.96m, respectively. This demonstrates that the path planning length using the method of this invention has a comprehensive advantage, with an average reduction of 5.1m. Furthermore, this invention is not only applicable to UAV path planning but can also be extended to path planning tasks in other fields such as mobile robots and autonomous driving, possessing broad application prospects.

[0045] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on its differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The above descriptions are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for optimal path planning of unmanned aerial vehicles (UAVs) combining neural networks, characterized in that, include: S1. Generate three-dimensional raster map data according to preset map parameters, and use the three-dimensional raster map data as samples for training the neural network module; S2. Update the heuristic function of the path planning module according to the cost increment output by the neural network module. The neural network module includes four input channels, which correspond to the obstacle and terrain cost, the target position, the current node position, and the next candidate node position, respectively. During the process of updating the heuristic function, the cost prediction function output by the neural network module corresponds to a 4-way parallel feature decoupling branch. Before S2, it also includes: a path planning module based on the A* algorithm that is associated with the cost increment, and the heuristic function of the path planning module is... , It is the cost estimate from the initial state to the target state via state n. It is the actual cost of going from the initial state to state n in the state space. It is the estimated cost of the optimal path from state n to the target state. The cost increment predicted by the convolutional neural network of the neural network module; The states in the heuristic function correspond to the position coordinates of the UAV, where the starting position of the UAV is... The drone's current location is , and These are the x-coordinate and y-coordinate values ​​of the starting position, respectively. and These are the x and y coordinates of the current position, respectively. S3. Generate the drone's flight path using the updated path planning module.

2. The method according to claim 1, characterized in that, S1 includes: Based on a preset raster 3D raster map dataset, generate 3D raster map data according to a preset size. The terrain features in the preset raster 3D raster map dataset include at least urban buildings, ruins, and bridges.

3. The method according to claim 1, characterized in that, In S2, the cost prediction function The corresponding four parallel feature decoupling branches include: the distance feature branch pointing to the fully connected layer outputting distance weights, the distance and height feature branch pointing to the fully connected layer outputting height weights, the steering feature branch pointing to the fully connected layer outputting steering weights, and the obstacle feature branch pointing to the fully connected layer outputting obstacle weights. This indicates the current location and status of the drone. This indicates the next node position status of the drone.

4. The method according to claim 3, characterized in that, Updated heuristics for: , , ,in, and Here are two modified cost functions. and Let state n and state n be the two states. The corresponding flight costs, The coordinates of the next target point, and These are the x-axis and y-axis coordinates of the next target point, respectively.

5. The method according to claim 4, characterized in that, The cost prediction function includes: , , ,in, Let x, y, and z be the coordinates of the node at the theoretical optimal distance, respectively. The coordinates of the nearest obstacle to the current position. , , These are the x-axis, y-axis, and z-axis coordinates of the nearest obstacle to the current position. For distance weighting function, This represents the maximum range of movement allowed in a single step. , and These are the x-axis, y-axis, and z-axis coordinates of the current position, respectively. The z-axis coordinate of the next target point. For the shift weight function, Let the obstacle weight function be... , , and These are the predicted cost increment values ​​corresponding to the distance weight function, height weight function, turning weight function, and obstacle weight function, respectively.

6. The method according to claim 3, characterized in that, Also includes: In the neural network module, a loss function L is set, where, , Based on the prediction of loss, To guide losses for safety, For smoothness regularization loss, For physical constraint loss, , , and They are respectively the corresponding , , and The weighting coefficients.

7. An optimal path planning system for unmanned aerial vehicles (UAVs) incorporating neural networks, characterized in that, include: Map generation module, neural network module, and path planning module; The map generation module is used to generate three-dimensional raster map data according to preset map parameters, and the three-dimensional raster map data is used as a sample for training the neural network module. The neural network module is used to update the heuristic function of the path planning module based on the cost increment output by the neural network module. The neural network module includes four input channels, which correspond to the obstacle and terrain cost, the target position, the current node position, and the next candidate node position, respectively. During the process of updating the heuristic function, the cost prediction function output by the neural network module corresponds to a 4-way parallel feature decoupling branch. Among them, the path planning module associated with the cost increment is built based on the A* algorithm, and the heuristic function of the path planning module is: , It is the cost estimate from the initial state to the target state via state n. It is the actual cost of going from the initial state to state n in the state space. It is the estimated cost of the optimal path from state n to the target state. The cost increment predicted by the convolutional neural network of the neural network module; The states in the heuristic function correspond to the position coordinates of the UAV, where the starting position of the UAV is... The drone's current location is , and These are the x-coordinate and y-coordinate values ​​of the starting position, respectively. and These are the x and y coordinates of the current position, respectively. The path planning module is used to generate the flight path of the UAV through the updated path planning module.

Images