Vision-based unmanned aerial vehicle airport landing area detection method

Abstract

The invention provides an unmanned aerial vehicle airport landing area detection method based on vision, and belongs to the technical field of unmanned aerial vehicle flight, and the method specifically comprises the steps: constructing a recognition model based on a YOLOv11 framework; training a recognition model by using the image marked with the airport information; inputting images acquired in real time into the trained identification model, and outputting detection results of the airport and the runway by the identification model; extracting a runway center line, marking the runway center line in a landing area, and designing an SCINet low-illumination image enhancement module in a backbone network of the recognition model to improve the recognition capability of the model in a low-illumination scene; and optimizing a feature extraction module of a backbone network and a neck network of the recognition model to enhance the perception capability of the feature extraction module to multi-scale features. And a runway center line extraction module is designed to assist the unmanned aerial vehicle to realize accurate runway alignment and autonomous landing.

Description

Technical Field

[0001] The present application relates to the field of drone flight, and in particular to a vision-based drone airport landing area detection method. Background Art

[0002] Autonomous landing is a crucial component of fixed-wing drones' routine flight missions. For fixed-wing drones operating at dedicated airports, where the runways are already relatively flat and safe, precise autonomous landing requires computer vision technology to accurately identify the airport and runway. This ensures the accuracy and safety of autonomous landings. With the successful implementation of this system, fixed-wing drones can significantly enhance their autonomy in routine flight missions, providing a solid and reliable safety guarantee.

[0003] In low-light conditions, traditional visual models encounter low contrast and significant noise in sensor-generated images, making feature extraction difficult. Airport runways, viewed from a drone's perspective, exhibit significant scale variations, with smaller objects appearing at long distances and richer details at close range. While the Feature Pyramid Network (FPN) of traditional convolutional neural networks (CNNs) can integrate multi-scale features, its perception of small objects is limited. Summary of the Invention

[0004] In view of this, the present application provides a vision-based drone airport landing area detection method, which solves the problems in the existing technology and improves the accuracy and real-time performance of drone landing area detection in complex environments.

[0005] This application provides a vision-based drone airport landing area detection method using the following technical solutions:

[0006] A vision-based method for detecting a drone's airport landing area includes the following steps:

[0007] Build a recognition model based on the YOLOv11 framework;

[0008] Use images labeled with airport information to train the recognition model;

[0009] The real-time collected images are input into the trained recognition model, which then outputs the detection results of the airport and runway.

[0010] Extract the runway centerline and mark it in the landing area;

[0011] The YOLOv11 framework includes a backbone network, a neck network, and a head network. During the construction of the recognition model, the SCINet low-light image enhancement module is added to the backbone network of the YOLOv11 framework. The SCINet low-light image enhancement module serves as the head input network of the recognition model. The SCINet low-light image enhancement module is used to increase image brightness and improve image quality in dark environments. The C3k2 module in the backbone network and neck network of the YOLOv11 framework is optimized to form a full-dimensional dynamic convolution C3k2_ODConv module. The full-dimensional dynamic convolution C3k2_ODConv module learns the complementary attention of the convolution kernel in four dimensions: the spatial size of the kernel space, the number of input channels, and the number of output channels of each convolution layer in a parallel manner, thereby performing fine-grained adjustments in the four dimensions.

[0012] The backbone network uses a convolutional neural network to convert raw image data into multi-scale feature maps and extract features from the input image at multiple scales;

[0013] The neck network constructs a more discriminative feature representation from the multi-scale features extracted by the backbone network;

[0014] The head network deeply processes the feature maps transmitted by the neck network, and ultimately outputs the corresponding class labels, accurate bounding boxes, and pixel-level segmentation masks of the objects in the image.

[0015] Optionally, the backbone network enhances the network's sensitivity to local detail features in the process of aggregating context information through a hierarchical aggregation context information module;

[0016] The hierarchical aggregation context information module includes a focus contextualization unit and a gated aggregation unit;

[0017] The focus contextualization unit utilizes a series of deep convolutional layers to encode visual context information at different scales;

[0018] The gating mechanism unit is used to control the flow of information, and the gating mechanism unit determines the transmission and discarding of information.

[0019] Optionally, the SCINet low-light image enhancement module uses a cascaded illumination learning process. The SCINet low-light image enhancement module consists of multiple stages, each of which estimates the illumination of the input image. Each stage further improves the image quality based on the previous stage, so that the SCINet low-light image enhancement module gradually optimizes the illumination characteristics of the image at different levels.

[0020] Optionally, the SCINet low-light image enhancement module introduces a weight sharing mechanism, and the parameters of the SCINet low-light image enhancement module in each stage are fully shared.

[0021] Optionally, the detection head of the head network adopts a decoupling structure to separate the target classification task from the bounding box regression task, and optimizes the target classification task and the bounding box regression task through independent branches.

[0022] Optionally, the overall classification task loss function of the recognition model is the sum of the cross entropy loss, the distribution focus loss function, and the bounding box regression loss function, and the overall segmentation task loss function of the recognition model is equal to the overall classification task loss function plus the segmentation loss.

[0023] Optionally, the method for extracting the runway centerline is:

[0024] Perform Canny edge detection on the binary mask image of the airport runway output by the head network to extract the edge lines of the airport runway;

[0025] Based on Hough transform, the edge lines of the extracted airport runway are detected to extract multiple straight line edges of the detection runway;

[0026] Extract the two longest straight lines from multiple straight line edges and whose actual parallelism meets the preset requirements as the road edge lines for runway detection;

[0027] The line connecting the intersection of the road edge lines detected by the two runways at the binary mask image and the centroid of the binary mask image is the road centerline.

[0028] In summary, this application has the following beneficial technical effects:

[0029] This application designs a SCINet low-light image enhancement module within the backbone network to improve the model's recognition capabilities in low-light scenarios. It also optimizes the feature extraction modules of the backbone and neck networks to enhance their ability to perceive multi-scale features. A runway centerline extraction module is designed to assist drones in achieving precise runway alignment and autonomous landing. This application can significantly enhance the model's environmental perception capabilities, assisting drones in achieving precise runway alignment and autonomous landing, thereby ensuring the safety of the landing process and providing strong support for fixed-wing drones performing airport landing missions. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0031] Figure 1 This is a flowchart of the vision-based drone airport landing area detection method for this application;

[0032] Figure 2 Identify the model framework diagram for this application;

[0033] Figure 3 This is a schematic diagram of the traditional dynamic convolution in the embodiment of the present application;

[0034] Figure 4 This is the schematic diagram of the full-dimensional dynamic convolution C3k2_ODConv module for this application;

[0035] Figure 5 A comparison diagram of the mechanisms of self-attention and focus modulation in the embodiments of the present application;

[0036] Figure 6 This is a schematic diagram of runway centerline extraction in an embodiment of the present application. DETAILED DESCRIPTION

[0037] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0038] The following describes the embodiments of the present application through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The present application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that, in the absence of conflict, the features in the following embodiments and embodiments can be combined with each other. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without making creative work are within the scope of protection of this application.

[0039] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on this application, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement an apparatus and / or practice a method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this apparatus and / or practice this method.

[0040] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present application. The illustrations only show components related to the present application and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0041] Additionally, in the following description, specific details are provided to provide a thorough understanding of the examples. However, one skilled in the art will appreciate that the aspects described can be practiced without these specific details.

[0042] The present invention provides a vision-based method for detecting a landing area of ​​an unmanned aerial vehicle (UAV) at an airport.

[0043] like Figure 1 and Figure 2 As shown in FIG, a vision-based method for detecting a drone airport landing area includes the following steps:

[0044] Build a recognition model based on the YOLOv11 framework.

[0045] The recognition model is trained using images labeled with airport information. The training images cover images under different lighting conditions, such as daytime, nighttime, dusk, and rainy weather, as well as images of drones at different heights and angles.

[0046] The images collected by the drone through the onboard camera in real time are input into the trained recognition model, and the recognition model outputs the airport and runway detection results; the images collected by the drone through the onboard camera in real time include visible light images and infrared images of the airport and runway.

[0047] Extract the runway centerline and mark the runway centerline in the landing area.

[0048] The YOLOv11 framework includes a backbone network, a neck network, and a head network. During the construction of the recognition model, the SCINet low-light image enhancement module is added to the backbone network of the YOLOv11 framework. The SCINet low-light image enhancement module serves as the head input network of the recognition model. The SCINet low-light image enhancement module is used to increase image brightness and improve image quality in dark environments. The C3k2 module in the backbone network and neck network of the YOLOv11 framework is optimized to form a full-dimensional dynamic convolution C3k2_ODConv module. The full-dimensional dynamic convolution C3k2_ODConv module learns the complementary attention of the convolution kernel in four dimensions: the spatial size of the kernel space, the number of input channels, and the number of output channels of each convolution layer in a parallel manner, thereby performing fine-grained adjustments in the four dimensions.

[0049] The backbone network uses a convolutional neural network to convert raw image data into multi-scale feature maps and extract features from the input image at multiple scales;

[0050] The neck network constructs a more discriminative feature representation from the multi-scale features extracted by the backbone network;

[0051] The head network deeply processes the feature maps transmitted by the neck network, and ultimately outputs the corresponding class labels, accurate bounding boxes and pixel-level segmentation masks of the objects in the image, realizing airport area detection and runway segmentation.

[0052] In this application, the backbone network serves as the main feature extractor, and a convolutional neural network is used to convert the original image data into multi-scale feature maps.

[0053] The neck component acts as an intermediate processing stage, using specialized layers to aggregate and enhance feature representations at different scales.

[0054] The head network component function is based on the deeply refined feature map, and through refined calculations, it simultaneously outputs the target positioning coordinates, category belonging probability and pixel-level segmentation mask.

[0055] During the model training phase, the output of this module will serve as a key parameter in the calculation of the loss function, and the network parameters will be continuously optimized through the back-propagation mechanism to achieve accurate prediction of target detection tasks.

[0056] The design of each network in this embodiment is as follows:

[0057] like Figure 2 As shown, the backbone network:

[0058] The backbone network plays a key role in feature extraction in object detection tasks, extracting features from the input image at multiple scales. This process involves stacking convolutional layers and specialized blocks to generate feature maps at different resolutions. Its architecture employs a layered design, including the SCINet low-light image enhancement module, the Conv convolution layer, the full-dimensional dynamic convolution C3k2_ODConv module, the hierarchical contextual information aggregation FocalMD module, and the C2PSA module.

[0059] The SCINet low-light image enhancement module processes raw images, improving image quality in low-light environments and boosting the performance of subsequent visual tasks. When a drone encounters inclement weather during landing, the visual features of the images captured by the sensor are weak and difficult to distinguish from the background, making it extremely difficult to detect landing areas. Existing detectors experience reduced detection performance in adverse environments. To address this issue, we introduce an image enhancement module into the object detection and segmentation network to improve model detection performance.

[0060] The Self-Correction Illumination Network (SCINet) is a deep learning framework designed specifically for low-light image enhancement. Its core goal is to optimize the illumination part of the image, thereby improving image quality in low-light environments and improving the performance of subsequent visual tasks, such as object detection and recognition.

[0061] The SCINet low-light image enhancement module consists of a self-correction module and an illumination estimation module. The output of the self-correction module is added to the original low-light input and serves as the input for the illumination estimation in the next stage. The SCINet low-light image enhancement module utilizes a cascaded illumination learning process. Composed of multiple stages, each stage estimates the illumination of the input image and further improves image quality based on the previous stage. This allows the SCINet low-light image enhancement module to progressively optimize the image's illumination characteristics at different levels. Furthermore, the SCINet low-light image enhancement module incorporates a weight sharing mechanism, fully sharing parameters across all stages. This reduces the number of model parameters and improves training and inference efficiency. Furthermore, the SCINet self-correction module ensures consistent results throughout the cascade, enhancing model stability and generalization. Furthermore, the network employs an unsupervised training loss, enabling training even in the absence of extensive labeled data and enhancing its adaptability.

[0062] Overall, the SCINet low-light image enhancement module performs well in low-light image enhancement tasks and can be combined with object detection models to significantly improve detection performance in low-light environments. In this solution, the SCINet low-light image enhancement module is used as the head input network of the YOLO backbone network. This design improves image quality in low-light environments while maintaining object detection performance, thereby enhancing the YOLO model's ability to recognize low-light targets.

[0063] The convolutional layer is the core component of the network, primarily responsible for feature extraction, parameter sharing, spatial structure preservation, and translation invariance. It extracts local features such as edges and textures through local receptive fields, and reduces parameters by sharing convolution kernels, improving computational efficiency and mitigating the risk of overfitting. The convolution operation preserves the spatial structure of the input data, enabling it to capture local patterns. It also reduces dimensionality through stride and pooling operations, enhancing feature abstraction.

[0064] One of YOLOv11's core innovations is the C3k2 module, which inherits the CSPNet architecture and optimizes gradient flow to improve feature reusability. Its design utilizes a combination of 1×1 and 3×3 convolution kernels, enhancing multi-scale feature extraction while reducing computational complexity. Compared to YOLOv8's C2f module, C3k2 further optimizes feature fusion, resulting in superior performance in small object detection tasks. This solution further improves upon the C3k2 module by combining it with full-dimensional dynamic convolution, enabling more precise feature extraction and enhancing the network's learning and representation capabilities.

[0065] Traditional dynamic convolution usually only achieves dynamics in the number of convolution kernels by weighted combination of multiple convolution kernels to adapt to different input features. The full-dimensional dynamic convolution C3k2_ODConv module further extends this strategy and adopts a multi-dimensional attention mechanism. This mechanism learns the complementary attention of the convolution kernel in four dimensions: the spatial size of the kernel space, the number of input channels, and the number of output channels of each convolution layer in parallel, thereby making fine-grained adjustments in four dimensions. The conventional convolution layer has a single static convolution kernel applied to all input samples. For the dynamic convolution layer, it uses a linear combination of n convolution kernels dynamically weighted with the attention mechanism, so that the convolution operation depends on the input. Mathematically, the dynamic convolution operation can be defined as:

[0066] y=(α w1 W1+…+α wn W n )*x;

[0067] in, are input features and output features, which have h×w c in and c out Channel, W i represents the i-th convolution kernel, is the attention scalar representation W i The weight of .

[0068] like Figure 3 and Figure 4 As shown, compared with the traditional dynamic convolution to calculate a single attention scalar Different from ODConv, ODConv utilizes a novel multi-dimensional attention mechanism to parallelly compute the attention of all four dimensions of the kernel space. and This strategy allows the network to process features in each dimension more efficiently while ensuring complementarity and synergy between dimensions, thereby enhancing the network's flexibility and adaptability. Therefore, this solution uses a full-dimensional dynamic convolution module in the model instead of traditional convolution operations to enhance the model's feature extraction capabilities.

[0069] While YOLOv11's pyramid pooling layer (SPPF) can reduce feature map size, improve computational efficiency, and enhance translation invariance, it can also lead to information loss, making the model insensitive to detailed features and reducing its ability to perceive positional information. To address this issue, this solution replaces the SPPF pyramid pooling layer with the Focal Modulation module, a hierarchical contextual information aggregation module. This enhances the network's sensitivity to local details and improves its perception of global structure during the contextual information aggregation process. SPPF stands for Spatial Pyramid Pooling Fast.

[0070] The fundamental principle of Focal Modulation is to use a focal modulation mechanism to capture long-range dependencies and contextual information in images. Traditional self-attention mechanisms require complex interactions and aggregations between each query token and other tokens to compute attention scores and capture contextual information. Focal Modulation, on the other hand, first aggregates spatial context into modulators to varying degrees, then injects these modulators into the query token in a query-dependent manner. Compared to self-attention, focal modulation simplifies the interactions and aggregation operations, making it more lightweight.

[0071] like Figure 5 This paper compares the mechanisms of self-attention and focused modulation in detail, demonstrating the context aggregation process in focused modulation. The left image in the figure below illustrates how the self-attention model generates outputs through the interaction between key k and query q, as well as the aggregation process. The center and right images illustrate how focused modulation first processes input features through a lightweight linear layer, followed by the selective integration of feature information using hierarchical context aggregation and gated aggregation mechanisms. Finally, the modulator interacts with the query q to generate output features.

[0072] The FocalMD module implements hierarchical contextual information aggregation through two modules: focal contextualization and gated aggregation. The focal contextualization mechanism uses a series of deep convolutional layers to encode visual contextual information at different scales. These layers capture visual features from near to far perspectives, ensuring that the neural network can deeply understand image information at different levels. This allows the network to maintain sensitivity to local details while aggregating contextual information and enhances its understanding of global structure.

[0073] The gating mechanism controls the flow of information. It determines which information should be passed and which should be discarded. In the FocalMD module, which aggregates contextual information hierarchically, the goal of gated aggregation is to selectively aggregate contextual information for each query token, enabling the network to determine which specific contextual information is critical for the current query token and focus on relevant information. This approach improves model efficiency and performance by reducing unnecessary information processing and enhancing attention to important features, thereby enhancing the model's object detection and image classification performance in complex and changing visual scenes.

[0074] The C2PSA module enables the network to focus on specific areas of interest by improving the SE attention mechanism, thereby improving the detection accuracy of objects of different sizes and locations.

[0075] Neck Network:

[0076] The neck network plays a central role in feature fusion and enhancement in object detection tasks. Its core goal is to construct more discriminative feature representations from the multi-scale features extracted by the backbone network, thereby significantly improving object detection performance. In this solution, the neck network is primarily based on the YOLOv11 neck network framework. This framework inherits the classic PAN-FPN architecture from the previous generation YOLO series neck network, cleverly combining the top-down feature fusion mechanism of FPN with the bottom-up information enhancement mechanism of PAN to construct an extremely efficient bidirectional feature pyramid structure. This architecture enhances the model's ability to detect objects at multiple scales. In actual operation, the neck network is responsible for fusing features at different scales and transmitting the fused features to the head network for object prediction. This process involves upsampling feature maps from different levels and concatenating features. Through these operations, the model efficiently captures multi-scale information, laying a solid foundation for accurate object detection.

[0077] Head network:

[0078] In the object detection and classification process, the head network is responsible for generating the final prediction results. It deeply processes the feature maps transmitted by the neck network, ultimately outputting the corresponding class labels, precise bounding boxes, and masks for fine-grained segmentation of objects in the image. In this solution, the head network is based on the YOLOv11 head network and has been specifically improved to achieve accurate detection of airport areas and instance segmentation of airport runways, meeting the object recognition needs of specific scenarios.

[0079] YOLOv11's detection head uses a decoupled structure, separating the target classification task from the bounding box regression task, and optimizing them through independent branches. This decoupling design effectively avoids feature confusion and improves the network's learning ability, allowing the classification and regression tasks to each focus on different feature extraction methods, thereby improving detection accuracy and convergence speed. In traditional coupled detection heads, classification information and position information share the same feature space, which may cause information interference and affect the final detection effect. However, YOLOv11 uses an independent branch structure, allowing the classification branch to focus on distinguishing target categories, while the regression branch more accurately fits the target's bounding box, ensuring more stable and reliable detection results.

[0080] The loss function in the recognition model of this application is designed as follows: the overall classification task loss function of the recognition model is the sum of the cross entropy loss, the distribution focus loss function and the bounding box regression loss function, and the overall segmentation task loss function of the recognition model is equal to the overall classification task loss function plus the segmentation loss.

[0081] The specific design of the loss function is:

[0082] In terms of loss function selection, the classification task uses the cross entropy loss function, the distribution focus loss function, and the bounding box regression loss function. First of all, the cross entropy loss function is a loss function commonly used in classification tasks. It can measure the accuracy of the network's predicted category and provide stable gradient updates, enabling the network to converge faster and improve classification performance. The cross entropy loss function is calculated as follows:

[0083]

[0084] Where S is the grid size, Indicates whether the i-th unit is included, p i (c) represents the probability that the target in the i-th grid cell belongs to category c in the prediction task. is the true label, indicating whether the target in the i-th grid cell really belongs to category c.

[0085] The bounding box regression task is mainly used to optimize the bounding box prediction accuracy in combination with the distribution focus loss function (DFL). It makes the prediction smoother and more stable by modeling the probability distribution of the target box coordinates, which is especially suitable for high-precision target detection. The distribution focus loss function (DFL) is calculated as follows:

[0086]

[0087] Where N is the number of samples; C is the number of categories; y ic is the true label of the i-th sample; p icis the predicted probability that the i-th sample belongs to category c; α is the balancing factor used to adjust the weights between positive and negative samples; γ is the focusing parameter used to control the degree of attention paid to difficult samples.

[0088] The bounding box regression loss function, bls, takes into account the overlap between the target bounding box and the ground-truth bounding box, and also incorporates center point offset and aspect ratio. This allows the network to converge faster during training while improving the detection of small objects and objects with large aspect ratio variations. This combination of loss functions effectively improves YOLOv11's adaptability to objects of varying scales and enables more accurate positioning of the detection box. The bounding box regression loss function, bls, is calculated as follows:

[0089]

[0090] Where, is the grid size; B is the number of bounding boxes predicted for each grid cell; Indicates whether the jth bounding box in the i-th grid cell is responsible for predicting the target; x, y are the coordinates of the center point of the bounding box. w, h are the width and height of the bounding box; λ coord Is the weight coefficient used to balance the losses of different parts. The overall classification task loss Loss c It can be obtained from the following formula, Loss c =cls+dfl+bls.

[0091] The segmentation task loss function adds a segmentation loss function mls based on the classification loss function. The segmentation loss still uses the binary cross entropy loss, specifically using the predicted segmentation area and the actual segmentation area to calculate the binary cross entropy loss pixel by pixel. The overall segmentation task loss is Loss m It can be obtained from the following formula, Loss m =cls+dfl+bls+mls.

[0092] To further improve computational efficiency and reduce computational overhead, YOLOv11 introduces two depthwise convolutions in the classification detection head, significantly reducing the number of parameters and computation while maintaining efficient feature extraction. This series of optimizations enables YOLOv11 to achieve better computational efficiency and faster inference speed while maintaining high accuracy, making it suitable for edge computing and real-time detection tasks.

[0093] This solution specifically designs a runway centerline extraction module for airport runway segmentation tasks. This module can accurately extract the runway boundary lines and centerlines based on the segmentation mask image, thereby supporting precise alignment and autonomous landing of drones. The method for extracting the runway centerline is as follows:

[0094] The binary mask image of the airport runway output by the head network is processed by Canny edge detection to extract the edge lines of the airport runway.

[0095] Based on Hough transform, the edge lines of the extracted airport runway are detected and multiple straight line edges of the detected runway are extracted.

[0096] Since the airport runway itself is long, straight and parallel on both sides, the two longest straight lines whose actual parallelism meets the preset requirements can be extracted from multiple straight line edges as the road edge lines for runway detection.

[0097] The line connecting the intersection of the road edge lines detected by the two runways at the binary mask image and the centroid of the binary mask image is the road centerline.

[0098] like Figure 6 As shown, Figure 6 The upper left is a binary mask detection map. Figure 6 The upper right picture shows the edge information extracted by Canny edge detection. Finally, after processing, the boundary line and center line of the runway can be obtained. Figure 6 Bottom middle picture.

[0099] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A vision-based method for detecting the landing area of ​​a drone airport, characterized in that: The steps include: Build a recognition model based on the YOLOv11 framework; Use images labeled with airport information to train the recognition model; The real-time collected images are input into the trained recognition model, which then outputs the detection results of the airport and runway. Extract the runway centerline and mark it in the landing area; The YOLOv11 framework includes a backbone network, a neck network, and a head network. During the recognition model construction process, an SC INet low-light image enhancement module is added to the backbone network of the YOLOv11 framework. The SCINet low-light image enhancement module serves as the head input network of the recognition model. The SC INet low-light image enhancement module is used to increase image brightness and improve image quality in low-light environments. The C3k2 module in the backbone network and neck network of the YOLOv11 framework is optimized to form a full-dimensional dynamic convolution C3k2_ODConv module. The full-dimensional dynamic convolution C3k2_ODConv module learns the complementary attention of the convolution kernel in four dimensions: the spatial size of the kernel space, the number of input channels, and the number of output channels of each convolution layer in a parallel manner, thereby performing fine-grained adjustments in four dimensions. The backbone network uses a convolutional neural network to convert raw image data into multi-scale feature maps and extract features from the input image at multiple scales; The neck network constructs a more discriminative feature representation from the multi-scale features extracted by the backbone network; The head network deeply processes the feature maps transmitted by the neck network, and ultimately outputs the corresponding class labels, accurate bounding boxes, and pixel-level segmentation masks of the objects in the image.

2. The vision-based drone airport landing area detection method according to claim 1 is characterized in that: The backbone network enhances the network's sensitivity to local detail features in the process of aggregating context information through a hierarchical aggregation context information module; The hierarchical aggregation context information module includes a focus contextualization unit and a gated aggregation unit; The focus contextualization unit utilizes a series of deep convolutional layers to encode visual context information at different scales; The gating mechanism unit is used to control the flow of information, and the gating mechanism unit determines the transmission and discarding of information.

3. The vision-based drone airport landing area detection method according to claim 1 is characterized in that: The SCINet low-light image enhancement module uses a cascaded illumination learning process. The SCINet low-light image enhancement module consists of multiple stages, each of which estimates the illumination of the input image. Each stage further improves the image quality based on the previous stage, so that the SCINet low-light image enhancement module gradually optimizes the illumination characteristics of the image at different levels.

4. The vision-based drone airport landing area detection method according to claim 3 is characterized in that: The SCINet low-light image enhancement module introduces a weight sharing mechanism, and the parameters of the SCINet low-light image enhancement module in each stage are fully shared.

5. The vision-based drone airport landing area detection method according to claim 1, characterized in that: The detection head of the head network adopts a decoupled structure to separate the target classification task from the bounding box regression task, and optimizes the target classification task and the bounding box regression task through independent branches.

6. The vision-based drone airport landing area detection method according to claim 1, characterized in that: The overall classification task loss function of the recognition model is the sum of the cross entropy loss, distribution focus loss function and bounding box regression loss function, and the overall segmentation task loss function of the recognition model is equal to the overall classification task loss function plus the segmentation loss.

7. The vision-based drone airport landing area detection method according to claim 1, characterized in that: The method to extract the runway centerline is: Perform Canny edge detection on the binary mask image of the airport runway output by the head network to extract the edge lines of the airport runway; Based on Hough transform, the edge lines of the extracted airport runway are detected to extract multiple straight line edges of the detection runway; Extract the two longest straight lines from multiple straight line edges and whose actual parallelism meets the preset requirements as the road edge lines for runway detection; The line connecting the intersection of the road edge lines detected by the two runways at the binary mask image and the centroid of the binary mask image is the road centerline.

Images