A method for dynamic re-planning of short-range flight path in an aircraft

Abstract

The application discloses a kind of short-range flight path dynamic re-planning method in aircraft: step 1, aircraft aerial image data is collected, and data set is obtained and divided;Step 2, target recognition network model is constructed based on YOLOv8 network;Step 3, the training set obtained in step 1 is used to train the target recognition network model constructed in step 2, and the trained target recognition network model is obtained;Step 4, using the trained target recognition network model obtained in step 3, real-time aerial image is processed, and the detected obstacle is output;Step 5, according to the result of step 4, real-time planning aircraft flight path.The application realizes the accurate identification and positioning of small target, while meeting the real-time requirement, so as to be able to make accurate judgment quickly in complex environment, reduce algorithm memory occupation, improve operation speed and algorithm reliability, suitable for short-range complex background, low contrast situation obstacle accurate and efficient identification.

Description

Technical Field

[0001] This invention belongs to the field of aircraft route planning technology, specifically relating to a method for dynamic replanning of short- and medium-range aircraft routes. Background Technology

[0002] In modern air-to-ground warfare, low-altitude and ultra-low-altitude penetration makes aircraft less detectable by ground radar due to terrain concealment and ground clutter, increasing their chances of completing missions. Therefore, low-altitude and ultra-low-altitude penetration offers significant advantages in modern warfare and is gradually becoming the primary method of penetration. Based on terrain features, terrain-following and terrain-avoidance technologies utilize navigation systems to find the most suitable flight path. These technologies greatly enhance an aircraft's penetration and survivability, with path planning being particularly crucial.

[0003] In modern warfare, battlefield environment awareness is crucial. Due to the complexity of the battlefield, the side that acquires battlefield information faster and more accurately often has a greater advantage. Therefore, the ability of aircraft to rapidly perceive dynamically changing battlefield environment information and to re-adjust and replan their missions is key to their success in modern warfare. Aircraft need to possess real-time online dynamic mid-range replanning capabilities. In short- and medium-range trajectory planning, there are high requirements for the real-time performance and computation speed of the algorithm. The large computational load brought about by the data raster structure of digital maps, as well as the complex constraints, especially the algorithm design problems caused by complex terrain constraints, result in a gap between the actual algorithm planning time and actual needs. When aircraft are performing trajectory planning, they also need to replan or plan in real time because the threat information is uncertain, incomplete, inaccurate, or dynamically changing. During the ground phase, the powerful computing capabilities and simulation methods of ground stations can generate optimal global reference tracks. However, when performing missions in the air, it is necessary to continuously adjust the flight path based on the current local terrain and real-time threats, dynamically calculate the flight track, and track the track to complete the flight mission and achieve effective threat avoidance. However, due to limited airborne computing resources, there is a serious contradiction between the speed and accuracy of short- and medium-range flight path planning for aircraft.

[0004] There are two problems:

[0005] 1. Obstacle threat perception:

[0006] (1) The target occupies few pixels. In aerial images, the target usually occupies few pixels. When using deep learning to detect targets, as the number of neural network layers increases, the semantic information of the feature map is continuously strengthened, but the receptive field also gradually increases, thus reducing the ability to detect small targets.

[0007] (2) The target is 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 prone to occlusion, which makes it impossible to extract complete feature information, seriously affecting the recognition accuracy.

[17] .

[0008] To address the challenges of detecting small targets and deformed occluded targets, the algorithm is optimized in terms of dataset augmentation, model construction, and convolutional kernel optimization.

[0009] 2. Dynamic trajectory planning

[0010] Through a survey and analysis of relevant research both domestically and internationally, it is evident that research on trajectory planning technology is quite comprehensive and yields abundant results. However, considering the requirements of actual mission scenarios, some shortcomings still exist:

[0011] (1) Most trajectory planning algorithms are based on two-dimensional space for search and planning, generating three-dimensional trajectories by limiting aircraft altitude, etc. However, the aircraft mission environment is complex and variable. In order to fully simulate the real mission environment, trajectory planning in three-dimensional space is required. Trajectory planning on a three-dimensional map will undoubtedly increase the algorithm's computation 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 failure to find the optimal solution.

[0013] (3) There is limited research on trajectory replanning algorithms. To address the real-time and reliability requirements of online real-time trajectory replanning algorithms for aircraft, improving algorithm speed and ensuring algorithm reliability can be achieved through appropriate redundancy design to guarantee flight safety for the aircraft and pilots.

[0014] (4) In terms of secondary optimization of generated path points, it is necessary to consider the dynamic constraints of the aircraft and environmental constraints to ensure that the generated trajectory is actually feasible. When applying optimal control methods 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, the solution time and the solution efficiency, and simplify the complex constraints.

[0015] Aircraft operating in complex and highly uncertain environments require rapid replanning of local flight paths in the event of sudden threats or changes in target location. Due to the real-time and safety requirements of missions, aircraft must react quickly and promptly upon detecting situations to complete flight path replanning. This necessitates fast and effective replanning algorithms with high success rates and safety requirements. Summary of the Invention

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

[0017] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

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

[0019] Step 1: Collect aerial image data from the aircraft, and obtain and divide the dataset;

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

[0021] Step 21: Add an output to the first C2f module in the backbone of the YOLOv8 network, construct 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. The feature map output from P3 of the YOLOv8 network is input to the Upsample module of the feature fusion module for upsampling. The upsampled result is then combined with the output of the first C2f module in the backbone network and concatenated in the Concat module of the feature fusion module. The concatenated result is then input to the c2f module of the feature fusion module. The first output of the c2f module is connected to two Conv modules added to the Head, and the second output 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 Concat module in the neck network.

[0023] Step 22, improve all the c2f modules in the network model obtained in step 21, specifically including the following operations: (1) replace the original ordinary convolution kernels in the Bottleneck of the C2f module with DCN convolution; (2) replace the original ordinary convolution kernels in the C2f module with DCN convolution.

[0024] Step 23, select the loss function;

[0025] Step 3: Use the training set obtained in Step 1 to train the target recognition network model constructed in Step 2, and obtain the trained target recognition network model.

[0026] Step 4: Using the trained target recognition network model obtained in Step 3, process the real-time aerial images 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: Collect aerial image data from the aircraft and label it with obstacles;

[0030] Step 12: Preprocess the image data labeled with obstacles to obtain a dataset;

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

[0032] Furthermore, in step 12, preprocessing refers to data augmentation of the image, including projective 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 for the model.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] This invention addresses the characteristics of small size and rapid change of aerial targets by constructing a target recognition network model for identifying small obstacle targets using the YOLOv8n model as the basic framework. This model incorporates a feature fusion module (such as...) within the neck network. Figure 2 (As shown in the green box) This invention adds two Conv modules to the Head module and replaces the original ordinary convolutional kernels in the c2f module and Bottleneck with DCN convolutions, achieving accurate identification and localization of small targets while meeting real-time requirements. This allows for rapid and accurate judgments in complex environments, reduces algorithm memory usage, and improves computational speed and algorithm reliability. It is suitable for accurate and efficient obstacle identification in long-distance, complex background, and low-contrast situations. Furthermore, addressing the issue of inconsistent quality in labeled datasets for aerial target detection and the scarcity of authoritative public datasets, this invention introduces WIoU as the model's bounding box loss function. This avoids excessive focus on excellent anchor boxes during model training, resulting in more accurate and efficient results.

[0036] In summary, the method of this invention enables an aircraft to plan a flight path that meets its performance requirements in the shortest possible time, enhances the aircraft's maneuverability in response to the surrounding situation, completes path planning within a limited time and space, completes strike or reconnaissance missions against various enemy military targets in short to medium ranges, minimizes the threat to the aircraft itself, increases mission success rate, and ensures aircraft safety. Attached Figure Description

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

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

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

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

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

[0042] Figure 6 It is a record of reasoning time for 13,500 experiments;

[0043] Figure 7 It refers to the inference speed after model optimization;

[0044] Figure 8 The detection results are under different environments;

[0045] Figure 9 It is the effect of trajectory planning;

[0046] Figure 10 It is a three-dimensional trajectory projection;

[0047] Figure 11 This is the effect of trajectory replanning in the joint simulation experiment.

[0048] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Detailed Implementation

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

[0050] Step 1: Collect aerial image data from the aircraft, obtain and divide the dataset, including the following sub-steps:

[0051] Step 11: Collect aerial image data from the aircraft and label it with obstacles;

[0052] Step 12: Preprocess the image data with obstacles marked in Step 11.

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

[0054] Step 13: Divide the preprocessed dataset obtained in Step 12 into a training set, a validation set, and a test set, specifically according to an 80-10-10 ratio.

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

[0056] Aircraft target detection differs from other target detection tasks because aerial targets are often relatively small and change rapidly, especially at long distances, against complex backgrounds, and in low-contrast conditions, making identification more difficult. Therefore, target recognition network models need to accurately identify and locate small targets while meeting real-time requirements to make rapid and accurate judgments in complex environments. Based on real-time requirements, the target recognition network model of this invention is built on the YOLOv8n model as its basic framework. 'n' is the model with the fewest parameters and the strongest real-time performance in the YOLOv8 series, suitable for deployment on airborne equipment with limited computing resources. Conventional YOLOv8 includes target detection networks with P5640 and P61280 resolutions and an instance segmentation model based on YOLACT, introducing more convolutional layers and residual modules to improve model accuracy and robustness. The YOLOv8 network consists of a backbone, a neck, and a head. Therefore, YOLOv8's path aggregation network has a three-layer structure, extracts features from three different scales, and allows information to flow between different scales, increasing the richness of the features extracted by the model and enabling the model to detect targets at different scales.

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

[0058] Step 21: To enhance the detection capability for small targets, this invention adds an output to the first C2f module in the backbone of the YOLOv8 network and constructs a feature fusion module (such as...) in the neck network. Figure 2 (As shown in the green box), and add two Conv modules to the 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 from the P3 of the YOLOv8 network is input to the Upsample module of the feature fusion module for upsampling. The upsampled result is then combined with the output of the first C2f module in the backbone network and fed into the Concat module of the feature fusion module for concatenation. The concatenation result is then fed into the c2f module of the feature fusion module. The first output of the c2f module is connected to the two Conv modules added to the Head, and the second output 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 upsampled result is concatenated with the output of the first C2f module in the backbone to obtain a 160x160 resolution feature map. Because the feature map output by the first C2f module has a higher resolution, fewer convolution and downsampling operations are performed compared to other output layers, resulting in greater preservation of small-scale information and stronger detection capability for small objects. Simultaneously, an Upsample module is constructed to preserve the semantic information of the lower layers, enabling the transfer 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 kernels in Bottleneck of the C2f module with DCN convolution, see Figure 3 (2) Replace the original ordinary convolution kernels in the C2f module with DCN convolution, see Figure 4 .

[0062] Step 23, Loss Function Optimization

[0063] To address the issue of inconsistent annotation quality in the aerial target detection datasets used in this invention, and the scarcity of authoritative publicly available datasets, the impact of varying annotation quality on the training process must be considered. Therefore, WIoU is introduced as the bounding box loss function for the model.

[0064] For aerial target detection applications, small targets occupying fewer pixels and occluded targets are common. However, during training, the model's loss function tends to assign greater weight to larger targets and high-quality annotations, which is detrimental to the target detection task in this invention. WIoU employs a dynamic, non-monotonic mechanism to evaluate anchor box quality, which avoids excessive focus on 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, and obtain the trained target recognition network model.

[0066] Step 4: Using the trained target recognition network model obtained in Step 3, 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] To verify the feasibility and effectiveness of the method of the present invention, simulation verification was performed below.

[0069] 1. Simulation System Architecture

[0070] like Figure 5As shown, the obstacle threat environment simulation system mainly consists of three parts: (1) a three-dimensional terrain database system with terrain scene simulation and obstacle threat simulation functions. Its main function is to construct a high-precision terrain environment scene in the three-dimensional environment visualization software, including mountains, hills, vegetation, sky environment, lighting environment, 24-hour changes, etc.; (2) a ground-based sensor simulation module including image recognition, matching, feature information extraction, etc. It obtains information through visible light and lidar sensors, processes the information to support the trajectory replanning module, and provides obstacle threat information; (3) the trajectory replanning module is based on a three-dimensional digital map generated by the fusion of digital elevation map and sensor information. It comprehensively considers factors such as obstacle threat, aircraft maneuverability constraints, and terrain to carry out trajectory replanning and realize a safe and efficient obstacle threat avoidance algorithm. This 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. This example uses the UE4 engine to build a 3D terrain database.

[0074] 2.2 Obstacle Model Establishment

[0075] By using 3DMAX software, common obstacles are modeled and textured to meet the requirements of the types and number of obstacles in the dataset.

[0076] 2.3 Module Information Transmission

[0077] The UDP (User Datagram Protocol) network transmission protocol is selected for information transmission between modules, and the validity of the data is ensured through message data verification.

[0078] 3. Simulation Experiment Analysis

[0079] 3.1 Simulation Environment

[0080] The experiment was run on a Windows 11 system with an AMD Ryzen 5600X CPU (3.7GHz base clock, 4.6GHz boost clock), 16GB DDR4 3600MHz RAM, an Nvidia GeForce RTX 4070 graphics card, CUDA version 12.2, CUDNN version 8.0, Python 3.9 as the programming language, PyTorch 2.1.1 as the deep learning framework, and CVXOPT library version 1.3.2 for solving convex optimization problems such as linear programming and quadratic programming.

[0081] 3.2 Simulation Experiment of Obstacle Threat Detection

[0082] 3.2.1 Dataset Establishment

[0083] To improve the effectiveness of the target detection algorithm, enhance its relevant metrics, and improve its versatility and robustness, this invention sets up relevant obstacle samples under four different weather conditions (cloudy, sunny, foggy, and rainy / snowy) and two terrain conditions (mountainous and plains). The obstacle types include five typical obstacles: radar, signal towers, factories, buildings, and power towers. The total sample set consists of 500 images.

[0084] 3.2.2 Hyperparameter Settings

[0085] The initial learning rate was set to 0.01, the initial momentum to 0.937, the weight decay factor to 0.0005, the batch size to 16, the number of training epochs to 200, the image input pixel size to 4k, and the IOU threshold to 0.7. The deformable convolution kernel was DCNV3, the convolution kernel operator compilation parameters were half-precision, and the CUDA version was 12.2.

[0086] 3.2.3 Evaluation Indicators

[0087] (1) Evaluation indicators of identification effect

[0088] The evaluation metrics used in this experiment were mAP@0.5 and mAP@0.5:0.95. Precision and recall were used as auxiliary metrics. The performance of the original YOLOv8n model and the improved model of this invention was compared in the simulation environment.

[0089] The model underwent 200 iterations of training, and the iteration with the best performance was selected to represent the final model result. Before the improvement, the highest mAP@0.5 value was 0.98, while after the improvement, the highest mAP@0.5 value was 0.985, an improvement of 0.51%. Before the improvement, the highest mAP@0.5:0.95 value was 0.60375, while after the improvement, the highest mAP@0.5:0.95 value was 0.61087, an improvement of 1.2%. The detection performance for power towers showed a significant improvement, increasing from 0.949 to 0.99, an improvement of 4.3%. This demonstrates the model's improved capabilities in detecting small and deformed targets, resulting in a substantial increase in target detection accuracy.

[0090] The improved model of this invention has slightly higher precision and recall than the original model.

[0091] (2) Identify speed evaluation indicators

[0092] To evaluate the real-time performance of the algorithm, 300 photos were randomly selected from the training set under different weather, scene, and lighting conditions. Each photo contained a varying number of obstacle samples. Simultaneously, to test the algorithm's robustness, 200 noisy images were randomly generated using random numbers that conformed to a Gaussian distribution. Therefore, the total number of test samples was 500. Inference was performed using a GPU, with initial settings default. The model's object detection process consisted of three stages: preprocessing, inference, and post-processing. After 500 experiments, the average time was calculated: 1.06ms for preprocessing, 16.87ms for inference, and 4.52ms for post-processing. It is evident that the longest-running stage was inference, as shown in the time log below. Figure 6 As shown.

[0093] To improve the real-time performance of the detection algorithm, the inference phase, which is the most time-consuming part of the detection process, is optimized. While the model can use full precision during training to improve its performance, the difference between using half-precision and full-precision in the inference phase has little impact on the inference results. Therefore, half-precision data such as FP16 and INT8 can be used for inference to reduce space usage and speed up inference. Furthermore, due to NVIDIA graphics cards' special optimization for half-precision data, the performance improvement on NVIDIA graphics cards is even more significant.

[0094] TensorRT is an inference acceleration library developed by NVIDIA for high-performance deep learning inference. It is optimized for NVIDIA GPUs, enabling fast and efficient deep learning inference in production environments and supporting various deep learning frameworks and hardware platforms. This section uses TensorRT to perform specific optimizations on the NVIDIA GPU architecture, optimizing operators, improving the computation graph flow, fusing specific computational operations to reduce computational load, and using FP16 half-precision inference. The model was optimized using TensorRT, and the same 500 experiments were conducted. 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% better than before optimization, meeting the real-time requirements of the algorithm in the scenario of this invention.

[0095] This invention addresses the common challenges in aerial photography, such as small targets, varied shooting angles, and severe object deformation. It improves small target detection and deformed target detection. Based on the improved YOLOv8n model, experimental verification was conducted in a simulation environment, simulating terrain and lighting conditions in actual flight missions. Simulations were performed in different scenarios to verify the algorithm's effectiveness. Actual detection results are as follows: Figure 8 As shown.

[0096] 3.3 Simulation Experiment of Track Replanning System

[0097] 3.3.1 Map Preprocessing

[0098] This invention selects a digital map of the Xiamen area, uses the bicubic Hermite interpolation algorithm to adjust the map accuracy, and uses the Osgeo third-party library to perform latitude and longitude coordinate transformation. Then, it generates a three-dimensional terrain map through merging, cropping, and projection.

[0099] 3.3.3 Global Track Planning

[0100] This invention employs the D*Lite algorithm for global trajectory planning, such as... Figure 9 As shown in the image. Randomly select the start and end points on the map, perform 3D global trajectory planning and replanning, and verify the algorithm's effectiveness, as shown below. Figure 10 As shown.

[0101] 3.3.4 Global Track Planning

[0102] This experiment uses the D*Lite algorithm for trajectory replanning. Two extreme cases, with obstacles of excessively large and small size, are selected to verify the algorithm's effectiveness. Under the small obstacle case, the replanning time is 0.4 seconds, memory usage is 8408 bytes, the number of search nodes is 339, and the trajectory length is 173.7 bytes, meeting the requirements for timeliness and memory usage. Furthermore, the planned trajectory conforms to helicopter dynamics requirements, and the trajectory length is close to the global optimum. Under the large obstacle case, the replanning time is 2.9 seconds, memory usage is 8408 bytes, the number of search nodes is 339, and the trajectory length is 176 bytes, again meeting the requirements for timeliness and memory usage. Simultaneously, the planned trajectory conforms to helicopter dynamics requirements, and the trajectory length is close to the global optimum.

[0103] 3.4 Joint Simulation Experiment

[0104] To verify the feasibility of the overall scheme, the terrain / obstacle threat environment system introduced above will be used to simulate the real aircraft mission environment. The obstacle threat detection simulation module and the trajectory replanning module will be used to provide flight trajectories for the aircraft model and conduct joint simulation verification.

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

[0106] After confirming that the information transmission function is working properly, the global track start coordinates are set to (117.6380, 24.6798) and the end coordinates are set to (120.5472, 23.2138), generating a global track. The track points are then transmitted to the terrain / obstacle threat environment system via UDP, and the system generates the corresponding global track points within the simulation environment.

[0107] During flight, the ground-following sensor simulation module senses the surrounding environment and sends the acquired image and distance information to the sensor simulation computer. The target detection algorithm of this invention is then used to identify obstacles and extract information.

[0108] The sensor simulation computer transmits the detected obstacle information to the trajectory planning computer. The trajectory replanning algorithm processes the obstacles identified by the method of this invention and updates the planning map, finally obtaining a new trajectory that avoids obstacles and satisfies aircraft dynamics.

[0109] Once the aircraft model receives information about the new waypoints, it will continue flying along the replanned waypoint according to its own flight control model.

[0110] After joint simulation testing, the aircraft trajectory replanning algorithm designed in this invention has the ability to detect typical ground 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 performance and reliability, and can achieve safe, real-time and reliable aircraft trajectory replanning.

Claims

1. A method for dynamic replanning of short- and medium-range flight paths of an aircraft, characterized in that, Specifically, the steps include the following: Step 1: Collect aerial image data from the aircraft, and obtain and divide the dataset; Step 2, constructing a target recognition network model based on the YOLOv8 network, includes the following sub-steps: Step 21: Add an output to the first C2f module in the backbone of the YOLOv8 network, construct a feature fusion module in the neck network, and add two Conv modules in the Head. The feature fusion module includes a Concat① module, a c2f module, an Upsample module, and a Conv module. The feature map output from P3 of the YOLOv8 network is input to the Upsample module of the feature fusion module for upsampling. The upsampled result is then combined with the output of the first C2f module in the backbone network and concatenated into the Concat① module of the feature fusion module. The concatenated result is then input to the c2f module of the feature fusion module. The first output of the c2f module is connected to two Conv modules added to the Head, and the second output 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, improve all the c2f modules in the network model obtained in step 21, specifically including the following operations: (1) replace the original ordinary convolution kernels in the Bottleneck of the C2f module with DCN convolution; (2) replace the original ordinary convolution kernels in the C2f module with DCN convolution. Step 23, select the loss function; Step 3: Use the training set obtained in Step 1 to train the target recognition network model constructed in Step 2, and obtain the trained target recognition network model. Step 4: Using the trained target recognition network model obtained in Step 3, process the real-time aerial images 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- and medium-range flight paths of an aircraft as described in claim 1, characterized in that, Step 1 includes the following sub-steps: Step 11: Collect aerial image data from the aircraft and label it with obstacles; Step 12: Preprocess the image data labeled with obstacles to obtain a dataset; Step 13: Divide the dataset into training set, validation set and test set.

3. The aircraft short-to-medium range trajectory dynamic replanning method as described in claim 2, characterized in that, In step 12, preprocessing refers to data augmentation of the image, including projective transformation, color transformation, introduction of random noise, image stitching, and cropping.

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

Images