Dynamic re-planning method for middle-short range flight path of airplane

Abstract

The invention discloses a dynamic re-planning method for a short and medium-range flight path of an airplane. The method comprises the following steps: step 1, acquiring aerial image data of the airplane, and obtaining and dividing a data set; step 2, constructing a target identification network model based on a YOLOv8 network; 3, training the target recognition network model constructed in the step 2 by adopting the training set obtained in the step 1 to obtain a trained target recognition network model; step 4, using the trained target identification network model obtained in the step 3 to process a real-time aerial image, and outputting a detected obstacle; and 5, planning the flight path of the airplane in real time according to the result of the step 4. According to the invention, the method achieves the precise recognition and positioning of a small target, meets the real-time requirements, can achieve the quick and accurate judgment in a complex environment, reduces the memory occupation of an algorithm, improves the calculation speed, improves the reliability of the algorithm, and is suitable for the precise and efficient recognition of an obstacle under the conditions of a short and medium distance complex background and low contrast.

Description

Technical Field

[0001] The invention belongs to the technical field of aircraft route planning, and in particular relates to a method for dynamic replanning of short- and medium-range aircraft tracks. Background Art

[0002] In modern air-to-ground combat, when an aircraft makes a low-altitude or ultra-low-altitude penetration, it is difficult for the aircraft to be detected by ground radar due to terrain shielding and interference from ground clutter, making it more likely to complete the combat mission. Therefore, low-altitude and ultra-low-altitude penetration has great advantages in modern warfare and has gradually become the main combat method for completing penetration in modern warfare. According to the terrain and topography, terrain following and terrain avoidance technology use the navigation system to find the most suitable flight path. These technologies have greatly improved the aircraft's penetration and survivability, among which track planning is particularly critical.

[0003] In modern warfare, the perception of the battlefield environment is the key to combat. Due to the complexity of the battlefield, the party that obtains battlefield information faster and more accurately often has a greater advantage. Therefore, the aircraft's rapid perception of dynamically changing battlefield environment information and its ability to re-adjust and re-plan tasks are the key to whether it can be competent in modern warfare. The aircraft needs to have real-time online dynamic mid-range re-planning capabilities. In medium- and short-range trajectory planning, there are high requirements for the real-time performance of the algorithm and the speed of algorithm solution; the large amount of calculation brought by the data grid structure of the digital map, as well as the complex constraints, especially the algorithm design problems brought by the complex terrain constraints, lead to a gap between the actual algorithm planning time and the actual needs; when the aircraft is planning the trajectory, it will also need to re-plan the trajectory or plan in real time because the threat information is uncertain, incomplete, inaccurate or dynamically changing. During the ground phase, the optimal global reference trajectory can be generated with the help of the powerful computing power and simulation methods of the ground station. However, when performing missions in the air, it is necessary to continuously correct the route according to the current local terrain and real-time threats, dynamically calculate the flight trajectory and track the trajectory to complete the flight mission and effectively avoid threats. However, due to limited onboard computing resources in the air, there is a serious contradiction between the speed and accuracy of the aircraft's medium- and short-range trajectory planning.

[0004] There are two aspects to the problem:

[0005] 1. Obstacle threat perception:

[0006] (1) The target occupies fewer pixels. In aerial images, the target usually occupies fewer pixels. When using deep learning to detect targets, as the number of neural network layers continues to increase, the semantic information of the feature map continues to increase, but the receptive field also gradually increases, so the detection ability of small targets decreases.

[0007] (2) The targets are severely deformed or occluded. Due to changes in flight altitude, flight attitude, and camera angle, the same target presents different features in the image. At the same time, the target is easily occluded, resulting in the inability to extract complete feature information, which seriously affects the recognition accuracy.

[17] .

[0008] In view of the difficulties in detecting small targets and deformed and occluded targets, the algorithm is optimized from aspects such as dataset enhancement, model construction, and convolution kernel optimization.

[0009] 2. Dynamic trajectory planning

[0010] Through the investigation and analysis of relevant research at home and abroad, the relevant research on trajectory planning technology is very sufficient and fruitful, but considering the requirements of actual mission scenarios, there are still some shortcomings:

[0011] (1) Most trajectory planning algorithms are based on searching and planning in two-dimensional space, and generate trajectories in three-dimensional space by limiting the aircraft's altitude. However, the aircraft mission environment is complex and changeable. In order to fully simulate the real mission environment, trajectory planning in three-dimensional space is required. Trajectory planning in three-dimensional maps will undoubtedly increase the algorithm's calculation time and space usage.

[0012] (2) In the trajectory replanning algorithm, due to the frequent changes in the map, the trajectory needs to be constantly updated, which makes it easier for the trajectory planning algorithm to produce local optimal solutions, resulting in a serious decrease in planning efficiency or the inability to obtain the optimal solution.

[0013] (3) There are few studies on trajectory replanning algorithms. In order to meet the real-time and reliability requirements of the aircraft online real-time trajectory replanning algorithm, improve the algorithm speed, ensure the algorithm reliability, and carry out appropriate redundancy design to ensure the flight safety of the aircraft and the pilot.

[0014] (4) In the secondary optimization of the generated path points, the dynamic constraints and environmental constraints of the aircraft need to be considered to ensure that the generated trajectory is actually feasible. When applying the optimal control method to solve the constraints and objective functions constructed based on the aircraft's dynamics, mission requirements and environmental constraints, it is necessary to consider the complexity of the model, solution time and solution efficiency, and simplify the complex constraints.

[0015] The flight space of an aircraft is complex and highly uncertain when performing a mission. In the event of a sudden threat or a change in the target's position, it is necessary to quickly replan the local track. Due to the real-time and safety requirements of the aircraft when performing a mission, the aircraft needs to handle the situation in a timely manner, respond quickly, and complete the track replanning work when it finds a situation. This requires the replanning algorithm to be fast and effective, and has high requirements for the success rate and safety of the planning. Summary of the invention

[0016] The purpose of the present invention is to provide a method for dynamic replanning of short-range and medium-range aircraft trajectories, so as to solve the problems of algorithm operation speed and reliability of existing aircraft trajectory planning.

[0017] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0018] A method for dynamic replanning of short-range and medium-range flight paths of an aircraft, comprising the following steps:

[0019] Step 1, collect aircraft aerial image data, obtain and divide the data set;

[0020] Step 2: Build a target recognition network model based on the YOLOv8 network, including the following sub-steps:

[0021] Step 21, add output to the first C2f module in the backbone network of the YOLOv8 network, build a feature fusion module in the neck network, and add two Conv modules in the Head;

[0022] The feature fusion module includes a Concat module, a c2f module, an Upsample module, and a Conv module; wherein the feature map output by P3 of the YOLOv8 network is input into the Upsample module of the feature fusion module for upsampling, and the upsampling result and the output of the first C2f module in the backbone network are input into the Concat module of the feature fusion module for splicing; the splicing result enters the c2f module of the feature fusion module, the first output end of the c2f module is connected to the two Conv modules added in the Head, and the second output end is connected to the Conv module of the feature fusion module; the output of the Conv module of the feature fusion module is connected to the Concat② module in the neck network;

[0023] Step 22, improving all c2f modules in the network model obtained in step 21, specifically including the following operations: (1) replacing the original common convolution kernel in the Bottleneck in the C2f module with the DCN convolution; (2) also replacing the original common convolution kernel in the C2f module with the DCN convolution;

[0024] Step 23, select the loss function;

[0025] Step 3, using the training set obtained in step 1 to train the target recognition network model constructed in step 2, to obtain a trained target recognition network model;

[0026] Step 4: Use the trained target recognition network model obtained in step 3 to process the real-time aerial image and output the detected obstacles;

[0027] Step 5: Plan the aircraft trajectory in real time based on the results of step 4.

[0028] Furthermore, step 1 includes the following sub-steps:

[0029] Step 11, collecting aircraft aerial image data and marking obstacles on it;

[0030] Step 12, preprocessing the image data with obstacles marked to obtain a data set;

[0031] Step 13: Divide the dataset into training set, validation set and test set.

[0032] Furthermore, in step 12, preprocessing refers to data enhancement of the image, including projection transformation, color transformation, introduction of random noise, image stitching and cropping.

[0033] Furthermore, in step 23, WIoU is used as the bounding box loss function of the model.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] Aiming at the characteristics of small size and fast change of aerial targets, the present invention uses the YOLOV8n model as the basic framework to construct a target recognition network model for identifying small target obstacles. Figure 2 Green box), and add two Conv modules in Head, and replace the original ordinary convolution kernel in the network model c2f module and Bottleneck with DCN convolution, so as to achieve accurate recognition and positioning of small targets, and meet the real-time requirements, so as to make accurate judgments quickly in complex environments, reduce algorithm memory usage, improve operation speed and algorithm reliability, and be suitable for accurate and efficient recognition of obstacles at long distances, complex backgrounds, and low contrast. At the same time, there is a situation where the quality of the dataset annotation for aerial target detection is uneven, and there are few authoritative public datasets. The present invention introduces WIoU as the bounding box loss function of the model to avoid excessive attention to excellent anchor boxes during model training, making the results more accurate and efficient.

[0036] In summary, the method of the present invention enables an aircraft to plan a track that meets the performance requirements of the aircraft in the shortest possible time, enhances the aircraft's maneuverability and response capability to the surrounding situation, completes path planning within a limited time and limited space, completes strike or reconnaissance missions against various enemy military targets in the medium and short range, and minimizes the threat to the aircraft itself, thereby minimizing the mission success rate and ensuring the safety of the aircraft. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a conventional YOLOv8 network structure;

[0038] Figure 2 It is the structure of the target recognition network model in the present invention;

[0039] Figure 3 It is the Bottleneck in the improved C2f module;

[0040] Figure 4 It is an improved C2f module;

[0041] Figure 5 It is the overall architecture of the obstacle threat simulation system;

[0042] Figure 6 It is a record of 13,500 experimental reasoning times;

[0043] Figure 7 is the inference speed after model optimization;

[0044] Figure 8 It is the detection effect under different environments;

[0045] Fig. 9 is the trajectory planning effect;

[0046] Fig.10 It is a three-dimensional track projection;

[0047] Fig.11 It is the trajectory replanning effect in the joint simulation experiment.

[0048] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. DETAILED DESCRIPTION

[0049] The present invention provides a method for dynamic replanning of short-range and medium-range flight paths for an aircraft, which specifically comprises the following steps:

[0050] Step 1, collecting aircraft aerial image data, obtaining and dividing the data set, includes the following sub-steps:

[0051] Step 11, collecting aircraft aerial image data and marking obstacles on it;

[0052] Step 12: preprocess the image data with obstacles marked obtained in step 11.

[0053] Preprocessing refers to data enhancement of images, including projection transformation, color transformation, introduction of random noise, image stitching and cropping.

[0054] Step 13, divide the preprocessed data set obtained in step 12 into a training set, a validation set, and a test set, specifically according to a ratio of 80-10-10.

[0055] Step 2: Build a target recognition network model based on the YOLOv8 network.

[0056] Because aircraft target detection is different from other target detection tasks, aerial targets are often relatively small and change faster, especially in long distance, complex background, and low contrast situations, identification is more difficult. Therefore, the target recognition network model needs to realize accurate identification and positioning of small targets, while meeting the real-time requirements, so that accurate judgments can be made quickly in complex environments. Based on the real-time requirements, the target recognition network model of the present invention is constructed with the YOLOv8n model as the basic framework, and n is the model with the least parameters and the strongest real-time performance in the YOLOv8 series model, which is suitable for deployment on airborne devices with limited computing resources. Conventional YOLOv8 includes a target detection network with P5640 and P61280 resolutions and an instance segmentation model based on YOLACT, which introduces more convolutional layers and residual modules, and can improve the accuracy and robustness of the model. The YOLOv8 network is composed of a backbone network (Backbone), a neck network (Neck) and a detection head (Head). Therefore, the path aggregation network of YOLOv8 has a three-layer structure, which extracts features from three different scales and flows information between different scales, increasing the richness of features extracted by the model and enabling the model to detect targets of different scales.

[0057] Step 2 includes the following sub-steps:

[0058] Step 21, in order to enhance the detection capability of small targets, in the present invention, the first C2f module in the backbone network (Backbone) of the YOLOv8 network is increased in output, and a feature fusion module (such as Figure 2 Green box), and add two Conv modules in Head;

[0059] The feature fusion module includes a Concat module, a c2f module, an Upsample module, and a Conv module; the 80x80 resolution feature map output by P3 of the YOLOv8 network is input into the Upsample module of the feature fusion module for upsampling, and the upsampling result and the output of the first C2f module in the backbone network (Backbone) are input into the Concat module of the feature fusion module for splicing; the splicing result enters the c2f module of the feature fusion module, the first output end of the c2f module is connected to the two Conv modules added in the Head, and the second output end is connected to the Conv module of the feature fusion module; the output of the Conv module of the feature fusion module is connected to the Concat② module in the neck network.

[0060] In this step, the upsampling result is concatenated with the output of the first C2f module in the backbone to obtain a feature map with a resolution of 160x160. Because the feature map output by the first C2f module has a higher resolution, fewer convolution and downsampling operations are performed than other output layers, more small-scale information is retained, and the detection ability for small targets is stronger. At the same time, in order to retain the semantic information of the lower layer, an Upsample module is constructed to realize the transmission of semantic information.

[0061] Step 22, improve all c2f modules in the network model obtained in step 21, specifically including the following operations: (1) replace the original ordinary convolution kernel in the Bottleneck in the C2f module with the DCN convolution, see Figure 3 ; (2) The original ordinary convolution kernel in the C2f module is also replaced by DCN convolution, see Figure 4 .

[0062] Step 23, loss function optimization

[0063] The quality of the dataset annotations for aerial target detection in this invention varies greatly, and there are few authoritative public datasets. The impact of different annotation qualities on the training process should be considered during training. Therefore, WIoU is introduced as the bounding box loss function of the model.

[0064] For aerial target detection applications, small targets that occupy fewer pixels and occluded targets are more common, but during the training process, the model's loss function will assign greater weights to larger targets and high-quality annotations, which is not conducive to the target detection task in this invention. WIoU uses a dynamic non-monotonic evaluation mechanism for anchor box quality, which can avoid excessive attention to excellent anchor boxes during model training.

[0065] Step 3: Use the training set obtained in step 1 to train the target recognition network model constructed in step 2 to obtain a trained target recognition network model.

[0066] Step 4: Use the trained target recognition network model obtained in step 3 to process the real-time aerial images and output the detected obstacles.

[0067] Step 5: Plan the aircraft trajectory in real time based on the results of step 4.

[0068] In order to verify the feasibility and effectiveness of the method of the present invention, the following simulation verification is carried out.

[0069] 1. Simulation system architecture

[0070] like Figure 5As shown, the obstacle threat environment simulation system mainly consists of three parts, namely: (1) a three-dimensional terrain database system with terrain scene simulation and obstacle threat simulation functions, the main function of which is to build a high-precision terrain environment scene in the three-dimensional environment vision software, including mountains, hills, vegetation, sky environment, lighting environment, 24-hour changes, etc.; (2) the ground sensor simulation module includes image recognition, matching, feature information extraction, etc., obtains information through visible light and laser radar sensors, processes the information to provide support for the track replanning module, and provides obstacle threat information; (3) the track replanning module is based on the three-dimensional digital map generated by the fusion of digital elevation map and sensor information, and comprehensively considers factors such as obstacle threats, aircraft maneuverability constraints, terrain, etc., to perform track replanning and realize a safe and efficient obstacle threat avoidance algorithm. The present invention uses this system as a simulation platform to simulate and verify the relevant algorithms.

[0071] 2. Subsystem design

[0072] 2.1 Three-dimensional terrain database system

[0073] Unreal Engine 4 (UE4) is a comprehensive game engine developed by Epic Games for creating high-quality games, virtual reality (VR), augmented reality (AR), and simulations. Here, a 3D terrain database is built based on the UE4 engine.

[0074] 2.2 Obstacle model establishment

[0075] By using 3DMAX software, common obstacles are modeled and mapped so that they can meet the requirements of obstacle types and quantities in the data set.

[0076] 2.3 Module Information Transmission

[0077] The UDP (User Datagram Protocol) network transmission protocol is selected to transmit information between modules, and the validity of the data is ensured by verifying the message data.

[0078] 3. Simulation experiment analysis

[0079] 3.1 Simulation Environment

[0080] The experiment was run on Windows 11 system, configured with AMD Ryzen 5600X CPU, main frequency 3.7 GHz, acceleration frequency 4.6 GHz, 16GB DDR4 3600 MHz running memory, Nvidia Geforce RTX 4070 graphics card, Compute Unified Device Architecture (CUDA) version 12.2, CUDNN version 8.0, programming language Python 3.9, deep learning framework Pytorch 2.1.1, and CVXOPT library version 1.3.2 for solving convex optimization problems such as linear programming and quadratic programming.

[0081] 3.2 Obstacle threat detection simulation experiment

[0082] 3.2.1 Dataset establishment

[0083] In order to improve the effectiveness of the target detection algorithm, enhance the algorithm's related indicators, and enhance the algorithm's versatility and robustness, the present invention sets relevant obstacle samples under four different weather conditions: cloudy, sunny, foggy, and rainy and snowy, and two terrain conditions: mountainous and plain. The obstacle types include five typical obstacles: radar, signal tower, factory, house, and electric tower. The total number of sample sets is 500 images.

[0084] 3.2.2 Hyperparameter Settings

[0085] The initial learning rate of the training is set to 0.01, the initial value of the momentum is set to 0.937, the weight decay coefficient is 0.0005, the batch size is set to 16, the training round is 200 times, the image input pixel is 4k, and the IOU threshold is set to 0.7. The deformable convolution kernel is DCNV3, the convolution kernel operator compilation parameter is half precision, and the CUDA version is 12.2.

[0086] 3.2.3 Evaluation indicators

[0087] (1) Recognition effect evaluation indicators

[0088] The evaluation indicators used in this experiment are mAP@0.5 and mAP@0.5:0.95. Precision and recall are auxiliary indicators. The effects of the original YOLOv8n model and the improved model of the present invention are compared in the simulation experiment environment.

[0089] The model was trained for 200 rounds, and the round with the best indicators was selected to represent the final model effect. The highest mAP@0.5 value before the improvement was 0.98, and the highest mAP@0.5 value after the improvement was 0.985, an increase of 0.51%. The highest mAP@0.5:0.95 value before the improvement was 0.60375, and the highest mAP@0.5:0.95 value after the improvement was 0.61087, an increase of 1.2%. Among them, the detection effect of the power tower category has been significantly improved, from the original 0.949 to 0.99, an increase of 4.3%. It can be seen that the model has improved its ability to detect small targets and deformed targets, and the accuracy of target detection has been greatly improved.

[0090] The precision and recall rate of the improved model of the present invention are slightly higher than those before the improvement.

[0091] (2) Recognition speed evaluation index

[0092] In order to evaluate the real-time effect of the algorithm, samples under different weather, scenes, and lighting conditions are randomly selected from the training set, with a total of 300 photos, each of which contains a varying number of obstacle samples. At the same time, in order to test the robustness of the algorithm, the present invention randomly generates 200 noise pictures. The noise pictures are generated by random numbers, and the generation of random numbers conforms to the Gaussian distribution. Therefore, the total number of test samples is 500, and GPU is used for inference. The initial setting value is the default parameter. The model divides the target detection of samples into three stages: pre-processing, inference, and post-processing. 500 experiments were performed to obtain the average value, and the average time for the pre-processing stage was 1.06ms, the average time for the inference stage was 16.87ms, and the average time for the post-processing stage was 4.52ms. It can be seen that the longest time-consuming stage is the inference stage, and the time record of this stage is as follows Figure 6 shown.

[0093] In order to improve the real-time performance of the detection algorithm, the inference phase, which takes the longest time in the detection, is optimized. The model can use full precision in the training phase to improve the performance and expression of the model, while the use of half-precision and full-precision in the inference phase has little effect on the model inference results. Therefore, half-precision data such as FP16 and INT8 can be used for inference to reduce space usage and speed up inference. At the same time, due to the special optimization of NVIDIA graphics cards for half-precision data, the improvement of inference performance on NVIDIA graphics cards is more obvious.

[0094] TensorRT is an inference acceleration library developed by NVIDIA for high-performance deep learning inference. It is optimized for NVIDIA GPUs, can achieve fast and efficient deep learning inference in production environments, and supports a variety of deep learning frameworks and hardware platforms. Here, TensorRT is used to optimize the model on the Nvidia graphics card architecture, optimize operators, improve the calculation graph process, fuse specific calculation operations to reduce the amount of calculation, and use FP16 half-precision inference. Use TensorRT to optimize the model and conduct the same 500 experiments. The experimental results are as follows: Figure 7 As shown by Figure 7 It can be seen that the average inference time of the optimized model is 11ms, which is 34.8% higher than before optimization, meeting the real-time requirements of the algorithm in the scenario of the present invention.

[0095] Aiming at the common problems of small shooting targets, many shooting angle changes, and serious deformation of the shooting objects in aerial photography, the present invention improves the detection of small targets and deformed targets. Based on the improved YOLOv8n model of the present invention, experimental verification is carried out in a simulation environment to simulate the terrain and lighting conditions in actual flight missions, and simulation verification is carried out in different scenes to verify the algorithm effect. The actual detection results are as follows: Figure 8 shown.

[0096] 3.3 Simulation experiment of trajectory replanning system

[0097] 3.3.1 Map Preprocessing

[0098] The present invention selects a digital map of Xiamen, uses a bicubic Hermite interpolation algorithm to adjust the map accuracy, converts longitude and latitude coordinates through an Osgeo third-party library, and then generates a three-dimensional topographic map through merging, clipping, and projection.

[0099] 3.3.3 Global trajectory planning

[0100] The present invention adopts D*Lite algorithm to perform global trajectory planning, such as Fig. 9 Randomly select the starting and ending points in the map, perform global planning and replanning of the three-dimensional trajectory, and verify the effect of the algorithm, as shown in Fig.10 shown.

[0101] 3.3.4 Global trajectory planning

[0102] This experiment uses the D*Lite algorithm for trajectory replanning. Two extreme cases of obstacles that are too large and too small are selected to verify the effect of the algorithm. In the case of small obstacles, the replanning time is 0.4s, the memory occupancy is 8408 bytes, the number of search nodes is 339, and the track length is 173.7, which meets the requirements of timeliness and memory occupancy. At the same time, the planned trajectory meets the requirements of helicopter dynamics, and the track length is close to the global optimal solution. In the case of large obstacles, the replanning time is 2.9s, the memory occupancy is 8408 bytes, the number of search nodes is 339, and the track length is 176, which meets the requirements of timeliness and memory occupancy. At the same time, the planned trajectory meets the requirements of helicopter dynamics, and the track length is close to the global optimal solution.

[0103] 3.4 Joint simulation experiment

[0104] In order to verify the feasibility of the overall plan, the terrain / obstacle threat environment system introduced in the previous article will be used to simulate the real aircraft mission environment, and the obstacle threat detection simulation module and the trajectory replanning module will be used to provide the aircraft model with a flight trajectory for joint simulation verification.

[0105] All hardware devices are turned on and running normally. Open the obstacle avoidance trajectory planning module software, set the aircraft parameters, and then establish a UDP information transmission connection between the modules. By entering the remote IP address and the port number of each module, ensure that information can be transmitted normally between the modules.

[0106] After testing that the information transmission function is normal, set the global track starting point longitude and latitude to (117.6380, 24.6798) and the end point longitude and latitude to (120.5472, 23.2138) to generate a global track. Then send the track points to the terrain / obstacle threat environment system via UDP transmission, and the system generates corresponding global track points in the simulation environment.

[0107] During the flight of the aircraft, the ground sensor simulation module senses the surrounding environment, sends the acquired image information and distance information to the sensor simulation computer, and uses the target detection algorithm of the present invention to identify obstacles and extract information.

[0108] The sensor simulation computer transmits the detected obstacle information to the track planning computer, and the track re-planning algorithm processes the obstacles identified by the method of the present invention and updates the planning map, and finally obtains a new track that avoids obstacles and satisfies aircraft dynamics.

[0109] When the aircraft model receives the information of the new track point, it will continue to fly along the re-planned track according to its own flight control model.

[0110] After joint simulation tests, the aircraft trajectory replanning algorithm designed in the present invention has the target detection capability of typical surface obstacles, and can output multi-source fusion information with obstacle type, size, and distance information. The trajectory replanning algorithm can make full use of known information and newly acquired information to replan the aircraft trajectory. The algorithm has real-time and reliability, and can realize safe, real-time and reliable aircraft trajectory replanning.

Claims

1. A method for dynamic replanning of short-range and medium-range flight paths of an aircraft, characterized in that: The specific steps include: Step 1, collect aircraft aerial image data, obtain and divide the data set; Step 2: Build a target recognition network model based on the YOLOv8 network, including the following sub-steps: Step 21, add output to the first C2f module in the backbone network of the YOLOv8 network, build a feature fusion module in the neck network, and add two Conv modules in the Head; The feature fusion module includes a Concat module, a c2f module, an Upsample module, and a Conv module; wherein the feature map output by P3 of the YOLOv8 network is input into the Upsample module of the feature fusion module for upsampling, and the upsampling result and the output of the first C2f module in the backbone network are input into the Concat module of the feature fusion module for splicing; the splicing result enters the c2f module of the feature fusion module, the first output end of the c2f module is connected to the two Conv modules added in the Head, and the second output end is connected to the Conv module of the feature fusion module; the output of the Conv module of the feature fusion module is connected to the Concat② module in the neck network; Step 22, improving all c2f modules in the network model obtained in step 21, specifically including the following operations: (1) replacing the original common convolution kernel in the Bottleneck in the C2f module with the DCN convolution; (2) also replacing the original common convolution kernel in the C2f module with the DCN convolution; Step 23, select the loss function; Step 3, using the training set obtained in step 1 to train the target recognition network model constructed in step 2, to obtain a trained target recognition network model; Step 4: Use the trained target recognition network model obtained in step 3 to process the real-time aerial image and output the detected obstacles; Step 5: Plan the aircraft trajectory in real time based on the results of step 4.

2. The method for dynamic replanning of short-range and medium-range flight paths of an aircraft as claimed in claim 1, characterized in that: Step 1 includes the following sub-steps: Step 11, collecting aircraft aerial image data and marking obstacles on it; Step 12, preprocessing the image data with obstacles marked to obtain a data set; Step 13: Divide the dataset into training set, validation set and test set.

3. The method for dynamic replanning of short-range and medium-range flight paths of an aircraft as claimed in claim 2, characterized in that: In step 12, preprocessing refers to data enhancement of the image, including projection transformation, color transformation, introduction of random noise, image stitching and cropping.

4. The method for dynamic replanning of short-range and medium-range flight paths of an aircraft as claimed in claim 1, characterized in that: In step 23, WIoU is used as the bounding box loss function of the model.

Images