Image-based airport runway foreign matter intelligent detection and clearance auxiliary processing system and method

Abstract

The invention relates to the technical field of airport safety monitoring, and discloses an image-based airport runway foreign matter intelligent detection and clearance auxiliary processing system and method. The acquisition module is used for acquiring original image data of a runway and a clearance area in real time through a high-definition camera; adopting an improved YOLOv11 target detection algorithm to identify birds, unmanned aerial vehicles or foreign matter targets in the clearance area; an improved YOLOv11 backbone is used as a main body feature extractor, and original image data is converted into a multi-scale feature map; the neck assembly combines feature maps of different scales and transmits the feature maps to the head assembly for prediction; the head assembly performs target prediction based on the reconstructed feature map; wherein a C2f block in the YOLOv11 neck assembly is replaced by a C3k2 block, and an intersection stage part and a space attention block are introduced; and a bird repelling or cleaning module is dispatched according to the detection result. According to the invention, the monitoring efficiency can be effectively improved, the detection precision is improved, clearance auxiliary processing is optimized, and the safety of the airport runway is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of airport security monitoring, and in particular, to an intelligent inspection and clearance assistance processing system and method for foreign objects on an airport runway based on images. Background Art

[0002] Modern airports have put forward higher requirements for flight safety. The collision between birds and aircraft is a major cause of flight accidents. It is a common phenomenon for birds to look for food and habitats near airports, especially near runways where bird activities are frequent, which easily leads to potential flight safety hazards.

[0003] Foreign objects on airport runways (such as small stones, screws, etc.) are called FOD (Foreign Object Debris), which pose a serious threat to aviation safety and may cause damage to aircraft tires, ingestion of foreign objects by engines, damage to the fuselage structure, or damage to the runway surface. Although runway foreign objects seem tiny, they may trigger a chain reaction, threatening the safety of aircraft, passengers, and airport operations. In addition, during the takeoff and landing of aircraft, the friction between the tread rubber of the tires and the ground generates high-temperature heat, which easily melts the tread rubber. The melted rubber will adhere to the surface of the runway light transmissive lens. After a long time, the oil and gas emitted by the aircraft and the micro-dust in the air will also adhere to the lens and form dirt, making the projected light dim and unable to meet the requirements of the standard light intensity, affecting the safe takeoff and landing of aircraft.

[0004] The existing detection of foreign objects on airport runways relies on manual inspections. Manual inspections require staff to drive vehicles or walk along the runway for inspections. Due to the large area of the runway, the inspection cycle is long. This results in the inability to detect foreign objects in a timely manner. For example, if a foreign object falls on the runway during the inspection interval, it may not be immediately noticed, thus increasing the risk of aircraft takeoff and landing. Some traditional detection technologies use fixed cameras with limited viewing angles, which are prone to monitoring blind spots. When foreign objects appear in the blind spots, they cannot be captured in a timely manner, affecting the comprehensiveness of detection.

[0005] In terms of detection accuracy, early image recognition algorithms, such as simple feature-matching-based algorithms, have poor recognition ability for foreign objects in complex backgrounds. The surrounding environment of airport runways is complex, with runway markings, lighting equipment, etc. Traditional algorithms are easily interfered by these background factors, misjudging the background as foreign objects or having difficulty accurately identifying small-sized foreign objects. Some early deep learning-based detection models, such as the early versions of the YOLO series of object recognition and localization algorithms based on deep neural networks, have relatively simple model structures and limited feature extraction capabilities. When facing foreign objects with various shapes and materials on airport runways, they cannot effectively extract the features of foreign objects, resulting in low detection accuracy.

[0006] In terms of clearance assistance processing: Most existing clearance processing systems are after-the-fact processing, that is, after a foreign object is discovered, relevant departments are manually notified for processing, lacking real-time and automation. There is a time difference between discovering a foreign object and processing it. During this period, the runway is still in an unsafe state.

[0007] There is a lack of intelligent assessment of foreign object risks. Different types of foreign objects have different degrees of impact on flight safety. Existing systems cannot quickly and accurately assess risks based on factors such as the characteristics and locations of foreign objects, resulting in unreasonable allocation of processing resources. It may lead to untimely processing of high-risk foreign objects and excessive resources invested in low-risk foreign objects.

[0008] Application and deficiencies of image recognition technology in airport runway foreign object detection and clearance processing:

[0009] (1) Background complexity: The backgrounds of airport runways and clearance areas are complex. There are various markings and textures on the runway, and there are background elements such as buildings and mountains in the clearance area. YOLOv10 may have difficulty accurately identifying foreign objects and clearance targets in such a complex background, and is prone to misjudgment. For example, some patterns of runway markings may be misidentified as foreign objects.

[0010] (2) Light change: The light conditions at the airport change greatly throughout the day, including low light in the early morning and evening, strong direct sunlight at noon, and the influence of runway lights at night. YOLOv11 is sensitive to light changes, and the recognition accuracy may drop significantly under different lights. For example, it is difficult to see small foreign objects under low light, and overexposure is likely to occur under strong light, resulting in the loss of targets.

[0011] (3) Detection of small foreign objects: Some foreign objects on the airport runway, such as small stones and screws, are relatively small in size and occupy a small proportion in the image. YOLOv10 may have problems with insufficient feature extraction when detecting small targets, resulting in missed detections.

[0012] (4) Clearance long-distance targets: For the clearance area, it is necessary to identify long-distance targets such as birds and drones. These targets are small in the image and may have blurred details due to the distance. YOLOv10 may not be able to accurately capture their features for effective recognition.

[0013] (5) Computational resource requirements: To ensure a high recognition accuracy, the model structure of YOLOv10 is relatively complex and the computational amount is large. In the real-time monitoring scenario of airport runways, a large amount of image data needs to be processed, which may not meet the real-time requirements, resulting in delayed detection results and the inability to detect foreign objects or clearance threats in a timely manner.

[0014] (6) Frame rate and accuracy coordination: When increasing the frame rate to meet real-time requirements, some recognition accuracy may be sacrificed; while pursuing high-precision recognition may lead to a decrease in the frame rate. How to find a balance between the two to ensure high recognition accuracy while meeting real-time requirements is a key issue to be solved.

[0015] (7) Diversity of foreign object types: There are a wide variety of foreign objects on the airport runway, including different materials such as metal, plastic, and rubber, with various shapes and colors. The pre-trained YOLOv10 model may not fully cover these foreign object features, and targeted optimization training is required to accurately identify various foreign objects.

[0016] (8) Characteristics of clear sky targets: The targets in the clear sky area, such as the flight postures of birds being diverse and the models of drones being various, have complex appearance features. YOLOv10 needs to be improved for these characteristics to enhance the recognition ability of clear sky targets.

[0017] (9) Data acquisition and annotation: It is difficult to obtain a large amount of high-quality image data containing foreign objects on the airport runway and clear sky targets and accurately annotate them. The quality and quantity of data directly affect the training effect of YOLOv10. How to efficiently acquire and annotate data is a fundamental issue for model training.

[0018] (10) Model update mechanism: As the airport environment changes and new foreign objects or clear sky targets appear, the model needs to be updated in a timely manner. Establishing an effective model update mechanism to enable YOLOv11 to quickly adapt to new situations and continuously maintain high recognition performance is a challenge faced in the long-term operation of this platform. Summary of the Invention

[0019] In view of the above problems, the purpose of the present invention is to provide an image-based intelligent inspection system and method for foreign objects on the airport runway and clear sky auxiliary processing, which can effectively improve the monitoring efficiency, enhance the detection accuracy, optimize the clear sky auxiliary processing, and improve the safety of the airport runway. The technical solutions are as follows:

[0020] The image-based intelligent inspection system for foreign objects on the airport runway and clear sky auxiliary processing includes:

[0021] 1) Image acquisition module: Used to obtain the original image data of the runway and the clear sky area in real time through a high-definition camera;

[0022] 2) Image processing module: Adopts the improved YOLOv11 (i.e., YOLOv11-m) target detection algorithm to identify birds, drones, or foreign object targets in the clear sky area;

[0023] The architecture of the improved YOLOv11 includes a backbone, a neck component, and a head component;

[0024] The backbone is the main feature extractor, which uses a convolutional neural network to convert the original image data into multi-scale feature maps; the neck component combines feature maps of different scales and transmits them to the head component for prediction; the head component makes object predictions based on the reconstructed feature maps; the improvement includes replacing the C2f block in the YOLOv11 neck component with a C3k2 block and introducing a cross-stage partial and spatial attention block to enhance the small object detection ability;

[0025] 3) Bird repelling module: used to repel birds by laser, ultrasonic wave or sound;

[0026] 4) Control center: used to generate a risk assessment report based on the detection results and dispatch the bird repelling module, and the risk assessment report calculates the risk level based on the bird species, density, flight trajectory and foreign object position; 5) Internet of Things communication module: used to transmit foreign object detection data and bird repelling status to the airport server in real time.

[0027] An intelligent inspection and clearance assistance processing method for foreign objects on the airport runway based on images, including the following steps:

[0028] Step 1: Image acquisition;

[0029] The original image data of the runway and clearance area is obtained in real time through high-definition cameras deployed on the airport runway and clearance area;

[0030] Step 2: Improve the main feature extractor and extract the main features;

[0031] The original image data is input into the main feature extractor, and the original image data is converted into multi-scale feature maps through a convolutional neural network; the main feature extractor is the backbone of the improved YOLOv11 architecture; the improvement includes: replacing the standard convolutions in the 3rd, 7th, and 11th layers of YOLOv11 with dynamic snake-shaped convolutions and embedding a collaborative attention mechanism in the C3 module;

[0032] Step 3: Reconstruct the neck component to complete intermediate processing;

[0033] Intermediate processing: First, replace the C2f block in the neck component of the YOLOv11 architecture with a C3k2 block, use multiple C3k2 blocks to sample and connect feature maps of different scales, aggregate and enhance the feature representations at different scales, and complete the reconstruction of the feature maps;

[0034] Step 4: Improve the head component to complete object positioning and output;

[0035] Object positioning and output: Introduce a new cross-stage part and a spatial attention block into the head component of the YOLOv11 architecture; Based on the original spatial pyramid pooling-fast block of the YOLOv11 architecture, as well as the newly introduced cross-stage part and spatial attention block, obtain the detection results of birds, drones or foreign object targets in the clear area based on the feature map reconstructed by the neck component;

[0036] Step 5: Establish a risk assessment model, calculate the threat level according to the detection results through the constructed threat level function, activate the bird repelling module to select the corresponding bird repelling strategy for bird repelling according to the threat level, or activate the cleaning drive module to clean foreign objects.

[0037] The beneficial effects of the present invention are:

[0038] 1) Improve detection efficiency: Based on the powerful image recognition speed and multi-scale feature fusion ability of YOLOv11-m, the present invention can quickly process a large amount of image data, and can process different features in the image in parallel, greatly shortening the processing time of a single image; And the multi-scale feature fusion enables the model to effectively identify targets at different resolutions, without the need for complex preprocessing and multi-stage processing of images. This enables the system to obtain and analyze the runway monitoring images in real time, detect them instantly or within a short time when foreign objects appear, greatly improving the detection efficiency, winning valuable time for the airport to take measures to deal with foreign objects in time, and reducing the risk of accidents caused by foreign objects staying on the runway for a long time.

[0039] 2) Improve detection accuracy: The present invention optimizes YOLOv11. Its improved network structure adds more combinations of convolutional layers and pooling layers, which can extract image features more deeply. At the same time, a more advanced feature extraction mechanism, such as the attention mechanism, is adopted, enabling the model to focus on the key features of foreign objects and ignore background interference. This enables foreign objects, whether small-sized or with a color similar to the background, to be accurately identified in a complex background, effectively reducing the false positive rate and false negative rate, and significantly improving the detection accuracy, providing more reliable foreign object detection results for the airport and ensuring flight safety.

[0040] 3) Optimize the airspace assistance processing: Based on the detection results of YOLOv11-m, the platform of the present invention constructs an intelligent risk assessment model. This model uses machine learning algorithms to quickly and accurately evaluate the risk level of foreign objects to flight safety according to various characteristics such as the size, material, and position of the foreign objects. At the same time, the platform has an automated linkage mechanism. According to the risk assessment results, it automatically notifies and allocates the corresponding processing departments and resources. For high-risk foreign objects, it can quickly initiate an emergency processing process and preferentially allocate resources for cleaning; for low-risk foreign objects, they can be processed according to the normal process. This not only realizes the automation and real-time nature of airspace processing, but also reasonably optimizes the allocation of processing resources, improves the efficiency and effect of airspace processing, and maximally guarantees the safe airspace state of the airport runway. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is an overall external shape effect diagram of the airport cleaning bird repellent vehicle.

[0042] Figure 2 It is a schematic diagram of the convolution process of the original image.

[0043] Figure 3 It is a working flow chart of the bird repellent system.

[0044] Figure 4 It is a working flow chart of the cleaning system.

[0045] In the figure: 1 - high-definition camera; 2 - laser emitter; 3 - pan-tilt; 4 - ultrasonic bird repellent; 5 - intelligent induced sound bird repellent cannon; 6 - brush head; 7 - dry powder sprayer; 8 - sprinkler; 9 - wheel; 10 - solar photovoltaic panel; 11 - airport cleaning bird repellent vehicle. SPECIFIC EMBODIMENTS

[0046] The following further describes the present invention in detail with reference to the drawings and specific embodiments.

[0047] The image-based intelligent foreign object detection and airspace assistance processing system for airport runways of the present invention includes: an image acquisition module, an image processing module, a bird repellent module, a control center, and an Internet of Things communication module. The airport cleaning bird repellent vehicle 11 of the present invention is as Figure 1 shown: The high-definition camera 1 is arranged on the top of the vehicle body. The bird repellent module includes a laser emitter arranged on the pan-tilt 3 on the top of the vehicle body, ultrasonic bird repellents 4 arranged on both sides of the vehicle body, and an intelligent induced sound bird repellent cannon 5; the brush head 6 is arranged at the bottom of the front end of the vehicle body, the dry powder sprayer 7 is arranged in the middle of the front end of the vehicle bottom, and two sprinklers 8 are respectively arranged on both sides of the dry powder sprayer 7; a solar photovoltaic panel 10 is also arranged on the top of the vehicle body.

[0048] (1) Image acquisition module:

[0049] The high-definition camera 1 deployed on the airport runway and the clearance area is used to obtain the original image data of the runway and the clearance area in real time.

[0050] 1) Real-time and accurate monitoring: Utilize the high-definition camera 1 to capture high-definition images of the airport ground and the airspace in real time, ensuring accurate monitoring of bird activities. The high-definition camera can capture high-definition images of the airport ground and the airspace in real time, which is crucial for monitoring bird activities. Through real-time image acquisition, bird activities can be quickly detected, and corresponding bird repelling measures can be taken.

[0051] 2) Environmental perception ability: In addition to bird monitoring, the high-definition camera can also sense other factors in the airport environment, such as weather, light, etc. These information are crucial for optimizing cleaning and bird repelling strategies. For example, under bad weather conditions, the cleaning plan and bird repelling methods can be adjusted according to the sensed weather information to ensure the cleanliness and safety of the airport.

[0052] 3) Intelligent decision-making support: The image data collected by the high-definition camera can provide strong support for the intelligent decision-making system. Through the analysis of the image data, it is possible to more accurately judge the areas and degrees that need to be cleaned, as well as what kind of bird repelling strategies to adopt.

[0053] At the same time, through the camera, airport staff can monitor the situation of birds on the airport runway and in the airspace in real time, which is conducive to promptly discovering abnormal situations and taking timely measures to deal with them, ensuring the ability to respond to emergencies when the decision is abnormal. Through the image information obtained by the camera, the current runway pollution situation and bird distribution situation can be judged, and corresponding cleaning and bird repelling decisions can be taken immediately, reasonably and effectively.

[0054] In this embodiment, the hardware configuration is as follows:

[0055] Use the Hikvision DS-2CD63 series 8-megapixel global shutter industrial cameras. The deployment spacing is calculated according to the ISO 34-2022 standard based on the runway width to ensure that the overlapping rate of the adjacent camera fields of view is ≥15%; the installation height is 3.5m ± 0.2m from the runway surface, and the tilt angle is 12°; equipped with polarizing lenses (such as Kenko PL-CIR 72mm) to suppress the reflection interference of the runway surface.

[0056] The dataset of this embodiment contains 12,850 labeled images (4,200 for birds / 3,150 for drones / 5,500 for foreign objects), covering 6 extreme conditions such as rain, fog, and night.

[0057] (2) Image processing module:

[0058] Adopt an improved YOLOv11 object detection algorithm to identify bird, drone or foreign object targets in the clearance area.

[0059] The architecture of the improved YOLOv11 includes a backbone, a neck component, and a head component; the backbone is the main feature extractor that uses a convolutional neural network to convert the original image data into multi-scale feature maps; the neck component combines feature maps of different scales and transmits them to the head component for prediction; the head component performs object prediction based on the reconstructed feature maps; the improvement includes replacing the C2f block in the YOLOv11 neck component with a C3k2 block and introducing a cross-stage partial and spatial attention block to enhance the small object detection ability.

[0060] This platform utilizes cutting-edge pose estimation and object detection technologies to improve the bird recognition rate by focusing on two main objectives: developing a comprehensive dataset for body condition classification based on bird pose data to improve the eviction operation, and using a deep learning model for pose key point localization to improve the recognition detection. Using the improved YOLOv11, a robust dataset was collected in different mining scenarios under low-light and high-light conditions. This dataset helps to extract detailed metrics such as bird actions and bounding box sizes for body condition classification. For recognition detection, this study adopted YOLOv11-m across public datasets and large datasets from Avibase and Birdsnap to ensure effective bird recognition in different scenarios. The implemented model demonstrated excellent performance in recognition detection and pose estimation, confirming the effectiveness of integrating the YOLOv11-m architecture into bird repellent image recognition. By strengthening the detection and analysis of bird postures and behavior compliance, this study significantly improved runway safety, potentially reducing the accident rate and enhancing the effectiveness of bird repellent operations.

[0061] Accurate long-range object recognition is crucial in airport bird eviction applications. Although recent advancements have improved real-time recognition, existing models, especially those focusing on monocular depth estimation, face accuracy challenges due to the limitations of supervised deep learning (DL). This study proposed a robust, real-time bird eviction object recognition system that utilizes time series and attention mechanisms to enhance depth estimation. Using RGB frames of depth maps from the KITTI and synthetic datasets, and a fine-tuned YOLOv11 model, our system achieved a root mean square error (RMSE) of 1.24 meters and an RMSE (log) of 0.18 meters for depth estimation, as well as object detection up to 5 meters in length. The model maintained high accuracy (96.4%), recall (93.67%), and F1 scores (93.33%) across different ranges, confirming the accuracy of YOLOv11-m, with an average inference time of 13 ms for short-range detection and 17 ms for long-range detection. These results highlight the potential of the system for deployment in bird repellent image recognition scenarios, outperforming existing models in terms of accuracy and computational efficiency.

[0062] When the YOLOv11 object detection model detects small objects in high-resolution images, its direct use has poor results. The solution to this problem is to segment the image into smaller blocks before inference. Subsequently, the cropped images (about 466 pieces) were exported to Roboflow for annotating bird categories. In Roboflow, the images were preprocessed, including automatic orientation and resizing (640×640 pixels). Subsequently, the following augmentation operations were applied in Roboflow: horizontal flipping, cropping 0-20%, rotation -15° to 15°, saturation -25% to +25%, and blurring (maximum blur radius of 2.5 pixels). A total of 2686 images were exported for model training.

[0063] After data augmentation, the dataset was divided into a training set (80%), a validation set (14%), and a test set (6%). The dataset was exported and used to train three versions of the YOLOv11 model: nano, small, and medium (using Ultralytics 8.3.13 in PyTorch 2.4.0). Training was carried out on a workstation computer equipped with an NVIDIA GeForce RTX 4070 Ti GPU. The YOLOv11 trainer ran for 30 epochs, with a batch size of 8 and an image size of 680. By examining the relationship between precision and recall when the confidence threshold changes, 0.4 was selected as the confidence threshold. To evaluate the performance of the three models (nano, small, and medium) on the validation set and test set, metrics including precision, recall, and F1-score were calculated using the following formulas: ; ; ;

[0064] where P represents precision, R represents recall, Tw represents correct pose estimation i.e., true positives, and Fw represents incorrect pose estimation i.e., false positives. True positives indicate that the prediction result matches the true label, false positives indicate incorrect predictions for positive samples, and false negatives mean that the model fails to predict the true label. In our training example, both the precision and recall for detecting various birds are relatively high. The performance of the three models is similar to that of the YOLOv11 medium model, and each metric is slightly higher, so they were used for subsequent testing and applications.

[0065] Table 1 Model Performance Metrics .

[0066] Various metrics, including precision, recall, and F1 score, showed that the three models performed similarly, with YOLOv11-m performing slightly better (Table 1). The model performed strongly in bird category recognition, correctly identifying 91% of birds. Recognition accuracy ranged from 0.85 to 0.90, and F1 scores ranged from 0.88 to 0.90. Our accuracy in detecting birds was comparable to another study, in which similar images were used to train an improved YOLOv8 model, achieving an accuracy of 0.86 for detecting birds. This suggests that the model has high accuracy and low misclassification rate in identifying birds, and has the potential for practical applications. Although other methods, such as color component analysis and Birdsnap classifiers, may report higher accuracy, the YOLOv11-m-based model has clear advantages in speed, efficiency, and automatic feature extraction.

[0067] To evaluate the effectiveness of YOLOv11-m in recognition under different backgrounds, we continue to conduct an in-depth evaluation of the model using various performance metrics such as precision, recall, F1 score, mean average precision (mAP), and average inference time.

[0068] Table 2 Recognition effectiveness of YOLOv11-m in different backgrounds .

[0069] As shown in Table 2, YOLOv11 maintains high performance in both short-range (50 meters) and long-range (more than 50 meters) detection. Specifically, the precision and recall of short-range detection are 96.40% and 93.67%, respectively. The F1 score of long-range detection (93.33%) is slightly lower than that of short-range detection (96.40%), but this drop is acceptable within the scope of real-time bird repellent applications.

[0070] In the short-range case, YOLOv11 achieves mAP values ​​of 96.40% at thresholds IoU=0.5 and IoU=0.75, while long-range detection achieves 93.29% and 93.28%, respectively. These results show that YOLOv11-m can maintain target localization accuracy even at higher IoU thresholds, which is an essential requirement in real-time bird repelling where accuracy is critical. In addition, as shown in Table 3, YOLOv11-m maintains robust performance under different environmental conditions. For example, in daylight conditions, the model achieves 96.8% accuracy, while low-light and nighttime settings produce accuracy values ​​of 92.2% and 88.4%, respectively. Although the performance drops slightly under adverse lighting, the model can still fully exert its performance due to its high integration level.

[0071] Table 3 Performance of YOLOv11-m in different environments .

[0072] Regarding efficiency, the YOLOv11-m model achieves a short-range detection frame rate of 30 FPS and a real-time long-range detection frame rate of 25 FPS, with an average latency of 35 ms. The Interference Time per Frame (ITPF) is 15 ms for short-range detection and 18 ms for long-range detection. These performance metrics highlight the suitability of YOLOv11-m for applications requiring fast response times. The combination of high precision, minimal latency, and efficient inference time enhances the potential of YOLOv11-m to achieve reliable real-time object detection during the bird deterrence process. The performance metrics of YOLOv11-m demonstrate its effectiveness in bird object detection and recognition, especially in real-time scenarios. YOLOv11-m has high precision and recall capabilities in both short-range and long-range detections, enabling the identification of relevant objects within close and far distances. The high mAP score of the model, particularly at a more stringent IoU threshold (0.75), highlights its precise object localization, which is crucial in environments for correctly identifying and tracking targets at different distances. Maintaining an average inference time of 13 ms - 17 ms per frame enables the model to meet the typical stringent real-time requirements in airport and runway applications. This performance further highlights the adaptability of the YOLOv11-m architecture, mainly through fine-tuning and integration with attention mechanisms, enabling it to operate reliably even in challenging environments. Additionally, the resilience of YOLOv11-m under various environmental conditions (daytime, low light, and adverse weather) demonstrates its potential versatility in field deployments. Although there is a slight performance degradation under low light and rainy conditions, the precision and recall metrics remain strong due to the model's integration of ConvLSTM and attention mechanisms (which help maintain detection quality despite noise caused by lighting or weather). These results indicate that YOLOv11-m can provide a scalable solution for real-time object recognition and depth assessment in fast-decision environments when adequately fine-tuned and paired with a depth estimation framework.

[0073] Visual task completion based on YOLOv11-m:

[0074] 1) Improve accuracy and reduce complexity: The YOLOv11-m variant achieves an excellent mean Average Precision (mAP) score on the COCO dataset while using 12% fewer parameters than its YOLOv11 counterpart, demonstrating improved computational efficiency without sacrificing accuracy.

[0075] 2) Resume task versatility: YOLOv11-m performs well in various bird deterrence applications, including pose estimation, object recognition, image classification, instance segmentation, and Oriented Bounding Box (OBB) detection.

[0076] 3) Optimized Speed and Performance: By redesigning the architecture and streamlining the training process, YOLOv11-m achieves faster processing speeds while maintaining a balanced efficiency between accuracy and computational power.

[0077] 4) Simplified Parameter Count: The reduction in parameters contributes to improved model performance without significantly affecting the overall accuracy of YOLOv11.

[0078] 5) Advanced Feature Extraction: YOLOv11-m has been improved in both the backbone and neck architectures, enhancing its feature extraction capabilities and enabling more precise object detection.

[0079] 6) Context Adaptability: YOLOv11-m demonstrates versatility in various deployment scenarios, including object detection, instance segmentation, pose estimation, and object-oriented detection.

[0080] The core of the YOLO architecture consists of three basic components.

[0081] First, the backbone is the main feature extractor that uses a convolutional neural network to convert raw image data into multi-scale feature maps. The backbone is a crucial part of the YOLO architecture and is responsible for extracting multi-scale features from the input image. This process involves stacking convolutional layers and specialized blocks to generate feature maps at various resolutions. The convolution process of the original image is as Figure 2 shown.

[0082] Second, the neck component acts as an intermediate processing stage. The neck combines features of different scales and transfers them to the head for prediction. This process typically involves upsampling and concatenating feature maps at different levels, enabling the model to capture multi-scale information effectively. Specialized layers are used for aggregation and enhancement of feature representations at different scales. YOLOv11-m maintains a structure similar to its predecessor, using initial convolutional layers to process the image. These layers form the basis of the feature extraction process, increasing the number of channels while gradually reducing the spatial dimensions. A notable improvement in YOLOv11-m is the introduction of the C3k2 block, which replaces the C2f block used in previous versions. The C3k2 block has a higher computational efficiency with the implementation of a cross-stage partial (CSP) bottleneck. It uses two smaller convolutions instead of one larger convolution, as in YOLOv8. The "k2" in C3k2 represents a smaller kernel size, which helps with faster processing while maintaining performance.

[0083] Third, the head component plays the role of prediction. Based on the reconstructed feature maps, it generates the final output for object localization and classification. Based on this established architecture, YOLOv11-m expands and enhances the foundation laid by YOLOv11, introducing such as Figure 2As shown, excellent detection performance is achieved through architectural innovation and parameter optimization. YOLOv11-m retains the Spatial Pyramid Pooling - Fast (SPPF) block from previous versions but introduces a new Cross Stage Partial with Spatial Attention (C2PSA) block afterwards. The C2PSA block is a notable addition that enhances spatial attention in the feature map. This spatial attention mechanism enables the model to more effectively focus on important regions in the image. By spatially pooling features, the C2PSA block allows YOLOv11-m to concentrate on specific regions of interest, potentially improving the detection accuracy for objects of different sizes and positions.

[0084] YOLOv11 brings significant changes by replacing the C2f block in the neck with the C3k2 block. The C3k2 module is designed to be faster and more efficient, improving the overall performance of the feature aggregation process. After upsampling and concatenation, the neck in YOLOv11-m adopts this improved block, thus enhancing speed and performance. A notable feature of YOLOv11-m is its greater focus on spatial attention through the C2PSA module. Paying attention to the spatial attention mechanism enables the model to focus on key regions in the image, potentially improving accuracy, especially for smaller or partially occluded objects.

[0085] YOLOv11-m effectively processes and reconstructs the feature map using multiple C3k2 blocks. This C3k2 block is placed in several paths in the head to handle multi-scale feature depths at different positions. The C3k2 block exhibits flexibility according to the value of the c3k parameter:

[0086] When c3k = False, the C3k2 module behaves similarly to the C2f block, utilizing a standard bottleneck structure;

[0087] When c3k = True, the bottleneck structure is replaced by the C3 module, which allows for deeper and more complex feature extraction.

[0088] Main features of the C3k2 block:

[0089] Faster processing speed: Compared to traditional convolutions, using two smaller convolutions can reduce the computational overhead of a single large convolution, resulting in faster feature extraction.

[0090] Parameter efficiency: C3k2 is a more compact version of the CSP bottleneck, making the architecture more efficient in terms of the number of trainable parameters.

[0091] Another notable addition is the C3k block, which provides enhanced flexibility by allowing a custom kernel size. This adaptability of C3k is particularly useful for extracting more detailed features from images, contributing to improved detection accuracy.

[0092] The head of YOLOv11 - m includes several CBS (Conv + Batch Normalization + Silu, convolutional layer + batch normalization + Silu activation function) layers after the C3k2 block. The delayer further refers to the feature map in the following ways:

[0093] • Extract relevant features for accurate object detection.

[0094] • Stabilize and standardize the data through batch normalization.

[0095] Utilize the Sigmoid linear unit (SiLU) activation function to achieve non - linearity, thereby improving the model performance. The CBS block is a basic component in both the feature extraction and detection processes, ensuring that the offset feature map is passed to the subsequent layers for bounding box and classification predictions.

[0096] Each detection branch of the final convolutional layer and the detection layer ends with a set of Conv2D layers, which reduce the features to the required number of output bounding box coordinates and class predictions. The final Detect layer integrates these predictions, including:

[0097] • Bounding box coordinates for locating objects in the image.

[0098] • Objectiveness scores indicating the presence of objects.

[0099] • Class scores for determining the classes of detected objects.

[0100] The three major applications of computer vision tasks are image recognition, object detection, and image segmentation.

[0101] The structures involved in the algorithm are shown in Table 4:

[0102] Table 4 YOLOv11 - m architecture 。

[0103] Due to the complexity of the airport environment, foreign objects on the airport runway have characteristics such as being tiny and difficult to detect, and having few scales for feature information extraction. Therefore, this platform adopts the YOLOv11 algorithm and strengthens small - target detection to complete the visual task. To improve the detection accuracy of airport foreign objects, the YOLOv11 algorithm used replaces the C2f block in the neck with a C3k2 block and pays more attention to spatial attention through the C2PSA module. This mechanism enables the platform to focus on the key areas in the image, thereby improving the accuracy, especially for smaller or partially occluded objects.

[0104] Algorithm features:

[0105] 1) High precision: The algorithm achieves high-precision image recognition and has a high detection completion rate for runway foreign object targets.

[0106] 2) Low complexity: The parameters used are 12% less than the counterparts of the popular YOLOv11 in the market, demonstrating a reduction in the complexity of system processing without sacrificing accuracy.

[0107] 3) Interconnectivity: The algorithm realizes the interconnection of visual tasks with pre-path selection and post-clearance processing. Through the internal segmentation of the detected image, it is assisted by forward negative feedback regulation and backward positive feedback regulation to complete real-time processing of platform work.

[0108] 4) Low latency: The completion of visual tasks is within 50 ms, achieving low latency, meeting the standards of airport clearance assistance, and realizing high-real-time system applications.

[0109] Steps of the image processing algorithm:

[0110] 1) Main feature extraction: The original image data is transformed into a multi-scale feature map after passing through the convolutional neural network, and the multi-scale feature map is input by the backbone, involving stacking convolutional layers and dedicated blocks to generate targets with various resolutions of features.

[0111] Completion: Extract relevant features for accurate target detection; stabilize and standardize the data through batch normalization.

[0112] Backbone network architecture of this embodiment:

[0113] Improvements based on YOLOv11-m include:

[0114] Replace the standard convolutions in the 3rd, 7th, and 11th layers with dynamic snake-shaped convolutions (DSConv);

[0115] Embed a collaborative attention mechanism in the C3 module.

[0116] Multi-scale feature generation:

[0117] The output feature map scales are 160×160 (shallow layer), 80×80 (middle layer), and 40×40 (deep layer), and the corresponding receptive fields are 56×56px, 112×112px, and 224×224px respectively.

[0118] 2) Intermediate processing: The neck combines features of different scales and transmits them to the head for prediction. Sample and connect feature maps of different levels, use dedicated layers for aggregation and enhancement of feature representations at different scales, enabling the system to capture more effective multi-scale information. This process is processed simultaneously using two smaller convolutions to achieve high efficiency.

[0119] The C3k2 block includes the following sub-modules:

[0120] Dynamic snake-shaped convolution: It is used to adapt to the geometric structure of the target. By dynamically adjusting the shape of the convolution kernel, it enhances the feature extraction ability for complex foreign object targets in the airport. By learning the offset of the input feature map, it dynamically adjusts the shape of the convolution kernel.

[0121] Collaborative attention convolution: It is used to improve the accuracy and robustness of detection. By introducing an attention mechanism, it enhances the feature expression in important regions of the feature map while suppressing unimportant regions, enabling the network to focus more on the key parts of the target.

[0122] Gaussian sampling convolution: It is used to reduce the impact of noise on the detection result to a certain extent and improve the stability of the system. By sampling the convolution kernel with a Gaussian distribution, it enhances the smoothness and noise resistance of feature extraction.

[0123] Funnel activation function: It is used to provide different gains in different activation intervals, enabling the network to better process input signals of different intensities. Through the special design of the non-linear activation function, it enhances the adaptability of the network to features of different intensities.

[0124] Feature fusion adopts a bidirectional feature pyramid strategy, simultaneously utilizing bottom-up and top-down feature information to improve the detection ability for targets of different scales and achieve the fusion and enhancement of multi-scale features.

[0125] When c3k = False, the behavior of the C3k2 module is similar to that of the C2f block, using a standard bottleneck structure.

[0126] When c3k = True, the bottleneck structure is replaced by the C3 module, which allows for deeper and more complex feature extraction.

[0127] The structural details of the C3k2 block are as Figure 2 shown.

[0128] Feature fusion adopts a bidirectional feature pyramid (BiFPN) strategy, and the weight coefficients are initialized with learnable parameters α = 0.3 and β = 0.7.

[0129] 3) Object localization and output: The Spatial Pyramid Pooling - Fast (SPPF) block and the Cross-Stage Partial with Spatial Attention (C2PSA) block achieve excellent detection performance through architectural innovation and parameter optimization, enabling the system to more effectively focus on important regions in the image. By spatially pooling features, the C2PSA block focuses on the set priority detection regions, thereby improving the detection accuracy for objects of different sizes and positions in the target region.

[0130] A. Design of the Spatial Attention Block (SAB):

[0131] The spatial attention block includes the following sub-modules: position-sensitive convolution and channel weight generator; the position-sensitive convolution is used to introduce position information in the convolution operation, enhance the perception ability of the target spatial position, and can better capture the features at different positions in the feature map by combining position encoding and convolution operation; the channel weight generator: is used to compress the number of channels of the input feature map through convolution according to the spatial weight; the structure includes:

[0132] Position-sensitive convolution (PSConv, 3×3 number of groups = 4);

[0133] Channel weight generator (the input feature map is first compressed to 1 / 4 of the number of channels through 1×1 convolution).

[0134] Spatial weight matrix calculation formula: ;

[0135] where, σ L is the LeakyReLU activation function, is the value of the intermediate feature at position (i, j).

[0136] B. Detection head output:

[0137] The output dimension of each detection head is: S×S×(5 + 2 + C), where: S is the number of grids (40 / 80 / 160), 5 is the target box parameters (x, y, w, h, conf), 2 is the bird repelling priority score, and C is the number of categories (birds / drones / foreign objects = 3 categories).

[0138] Object localization is used to locate the bounding box coordinates of the object in the image, specifically as follows:

[0139] (a) Detection head architecture configuration:

[0140] CBS layer topology:

[0141] Each detection branch contains 3 cascaded CBS blocks, and the parameters are as follows:

[0142] CBS (in_c = 256, out_c = 256, k = 3, s = 1) # Feature-preserving convolution;

[0143] CBS (in_c = 256, out_c = 128, k = 1, s = 1) # Channel compression;

[0144] CBS (in_c = 128, out_c = 256, k = 3, s = 1) # Spatial feature enhancement;

[0145] A cross-stage residual connection (refer to the Res2Net structure) is inserted after each layer, and the skip connection weight α = 0.2;

[0146] Batch normalization implementation:

[0147] Adopt dynamic momentum BN (DynBN), and the momentum coefficient changes with the number of training epochs: (t is the current epoch number, with an upper limit of 0.999).

[0148] (b) Specific process of coordinate regression:

[0149] Feature map mapping to anchor boxes:

[0150] Perform grid division on the input feature map (40×40×256), and each grid point is associated with 3 prior anchor boxes (sizes obtained through K-means clustering):

[0151] Small target anchor box: 12×16px; Medium target anchor box: 32×40px; Large target anchor box: 68×80px.

[0152] Convolution output dimension calculation:

[0153] Number of output channels for each grid point: 3×(5 + num_classes); (where 5 = 4 coordinate offsets + 1 confidence, num_classes = 3).

[0154] Coordinate non-linear decoding formula: ; ; ; ; Among them, t x , t y , t w and t h are the original network outputs after being activated by Silu; c x and c y are the coordinates of the top-leftmost grid point of the feature map; w a and h a are the reference sizes of the anchor boxes; b x is the abscissa of the center point of the bounding box, b y is the ordinate of the center point of the bounding box, b w is the width of the bounding box, b h is the height of the bounding box; is the Sigmoid function (limiting the output to 0 - 1); A learnable scale coefficient λ = 1.05 (initial value) is introduced for the dynamic scaling factor and updated through gradient descent: ;

[0155] Design of Loss Function in Training Phase:

[0156] Positive Sample Matching Strategy:

[0157] Adopt Task-Aligned Assigner: Calculate Alignment Metric: ( is the classification integral, is the IoU between the predicted box and the GT);

[0158] Select the top-k (k = 10) samples as positive samples.

[0159] Table 5 Experimental Data Support .

[0160] According to the experimental data in Table 5, compared with the traditional linear decoding ( ), the non-linear decoding formula of the present invention reduces the small target positioning error by 45.2%; compared with the fixed anchor box, the dynamic scaling factor of the present invention is optimized by gradient descent, and the mAP@0.5 is increased by 7.7%; the decoding delay is reduced by 54.2%.

[0161] (3) Bird Repelling Module:

[0162] Two laser emitters 2 under the high-definition camera of the intelligent bird repeller are fixed on the pan-tilt 3. The internal laser generating device emits green laser outward. At the same time, the pan-tilt 3 can rotate 360° in the horizontal plane, and the emitter can rotate 360° in the vertical plane. With the image recognition of the system, it realizes all-round accurate shooting at birds, achieves a high-efficiency bird repelling effect, and realizes point-to-point accurate bird repelling.

[0163] Used for bird repelling by laser, ultrasonic wave or sound; the bird repelling module includes a laser bird repelling unit, an ultrasonic bird repelling unit and a sound bird repelling unit; the laser bird repelling unit emits a green laser beam with a wavelength of 532nm, and the power range is 500mW - 5W, and the scanning speed is dynamically adjusted according to the bird density; the ultrasonic bird repelling unit operates at a frequency of 12 kHz - 25kHz and adopts a frequency sweeping mode to extend the bird adaptation cycle; the sound bird repelling unit stores more than 2000 kinds of bird natural enemy calls and screams and supports a random playback strategy.

[0164] 1) Laser Bird Repelling:

[0165] Specific Parameters in Laser Bird Repelling:

[0166] Wavelength: Laser bird repelling usually uses a green laser with a wavelength of 532nm.

[0167] Power: The power used by laser bird repellent devices typically ranges from several hundred milliwatts (mW) to several watts (W).

[0168] In practical applications, the laser bird repellent emits a laser beam with a certain wavelength and power, simulating the conditions for biological visual reflection, causing birds to feel threatened and fly away.

[0169] Algorithm steps:

[0170] A. Laser beam generation: The laser bird repellent device first generates a green laser beam with a wavelength of 532 nm and a power of 500 mW. This wavelength of laser beam is highly sensitive to the avian visual system and can effectively attract the attention of birds.

[0171] B. Laser beam scanning: The laser beam is scanned within the target area at a certain speed and angle through a precise aiming and scanning system. During the scanning process, the laser beam forms a "rod-shaped" green laser with a diameter of 152 mm, simulating the conditions for biological visual reflection.

[0172] C. Bird perception and repulsion: When birds perceive the laser beam, they will receive a strong visual stimulus, as if seeing a large green rod. This stimulus will make birds feel uneasy and uncomfortable, and thus instinctively fly away from the target area, achieving the effect of bird repulsion.

[0173] D. Real-time monitoring and adjustment: The laser bird repellent device is usually equipped with a high-definition camera and intelligent recognition algorithms, which can monitor the activities of birds in the target area in real time. When birds are detected, the device will quickly activate the laser beam for repulsion. At the same time, the bird control personnel can also remotely control the device to adjust the parameters of the laser beam to adapt to the bird activities in different scenarios.

[0174] 2) Ultrasonic bird repellent:

[0175] Designing the ultrasonic bird repellent 4 into an integrated circular shape can achieve the maximum effect. As Figure 1 shown, it can enhance the signal transmission. The circular design can ensure that the ultrasonic bird repellent device covers a larger area, thus more effectively repelling birds. And it makes the ultrasonic bird repellent device easier to install in various environments, such as farmlands, gardens or around buildings. It can also reduce the direct aggression towards birds, making it easier for birds to accept when receiving the ultrasonic signal and reducing unnecessary harm. In addition, the circular design can improve the stability and durability of the device, further increasing its service life and reliability.

[0176] The frequency range commonly used by the ultrasonic bird repellent unit is 12 KHz - 25 KHz. Ultrasonic waves within this range have a repellent effect on birds because they are sensitive to ultrasonic waves in this frequency range.

[0177] The working modes of ultrasonic bird repellers usually include two types: frequency sweeping and fixed frequency. Frequency sweeping means continuously changing the frequency points within a certain frequency range and working in a cycle, while fixed frequency means randomly selecting a frequency point and continuously working for a period of time.

[0178] The ultrasonic bird repelling technology interferes with the normal activities of birds by emitting ultrasonic waves of specific frequencies, thereby achieving the purpose of driving away birds. Its working principle involves the propagation and attenuation characteristics of ultrasonic waves in the medium, and these characteristics can be described by the above mathematical formulas. In addition, the frequency range and working mode of the ultrasonic bird repeller are also important factors to be considered in its design.

[0179] 3) Sound bird repelling:

[0180] To achieve the maximum bird repelling effect, the intelligent induction sound repelling cannon 5 is adopted, which can control bird repelling more precisely and efficiently. Adjust the sound output according to environmental conditions to better attract the attention of birds and make them leave the target area. At the same time, using the intelligent induction sound repelling cannon instead of physical attack methods can significantly reduce the risk of harming birds and reduce the impact on surrounding human activities. The design of the intelligent induction sound repelling cannon aims to drive away birds rather than harm them, which helps to avoid unnecessary damage or interference to bird habitats. Using sound repelling technology is more low-key than other methods, reducing interference to the surrounding people and protecting their privacy.

[0181] In the part of sound bird repelling, in view of the deficiencies of the existing technology, the following improvements are mainly made:

[0182] A. Enhancement of the richness and realism of sound types:

[0183] Richness of sound types: In the existing technology, sound bird repellers often use several fixed sounds for repelling, which easily leads to birds developing adaptability. In the improved solution, by adopting the latest digital voice chip repository technology, it can store more than 2000 kinds of sounds, greatly increasing the richness of sound types. These sounds are not limited to the calls or wails of birds' natural enemies, but can also include various sounds in nature, and even high-fidelity sound waves of specific frequencies, thus improving the diversity and effectiveness of bird repelling.

[0184] B. Optimization of the sound playback strategy:

[0185] Random playback and frequency adjustment: In the existing technology, sound bird repellers often use fixed playback sequences and frequencies, which are easily adapted by birds. In the improved solution, through random playback and frequency adjustment, it is difficult for birds to predict the occurrence pattern of sounds, thereby extending their adaptation period. For example, the ultrasonic bird repeller uses single-chip microcomputer technology to achieve segmented operation, randomly emitting ultrasonic waves of a certain frequency within a certain frequency range each time to extend the adaptation period of birds.

[0186] Timing and volume adjustment: The improved solution also supports computer clock timing to avoid delays in work. At the same time, the volume can be adjusted as needed to ensure that while driving away birds, it does not interfere with humans and other creatures. According to the activity patterns of birds and the operational needs of the airport, the sound playback time and interval are intelligently scheduled. By setting a timed playback task, it is ensured that the birds are effectively driven away during peak activity periods or key areas.

[0187] (4) Control center and IoT communication module:

[0188] The control center is used to generate a risk assessment report based on the detection results and dispatch the bird repelling module. The risk assessment report calculates the risk level based on the bird species, density, flight trajectory and foreign object location. The IoT communication module is used to transmit foreign object detection data and bird repelling status to the airport server in real time. The hierarchical response mechanism is shown in Table 6, and the bird repelling system workflow is shown in Figure 3 shown.

[0189] 1) Multi-dimensional threat feature extraction:

[0190] Dynamic target feature library construction:

[0191] Input the detection results of YOLOv11-m (target category, position, size, speed) and construct the feature vector in time series: ;

[0192] Added motion trajectory prediction: Kalman filtering (process noise Q=0.1I4×4, observation noise R=0.5I2×2) is used to estimate the position in the next 3 seconds.

[0193] Environmental status coding: Real-time integration of meteorological data (visibility, wind speed) and flight schedules (ADS-B signals) of the airport digital twin system.

[0194] 2) Risk Assessment Model:

[0195] Construct a threat function: ;

[0196] Weight coefficients: α=0.4, β=0.3, γ=0.2, δ=0.1 (through regression fitting of airport historical accident data).

[0197] Normalization factor: (The benchmark value corresponds to birds with a wingspan ≥ 1m); ; (d is the distance from the center line of the runway, in meters).

[0198] 3) Bird repelling strategy selection:

[0199] Table 6 Hierarchical response mechanism 。

[0200] The strategy selection analysis is as follows:

[0201] A. Evaluation of single bird repelling method:

[0202] Laser bird repelling: Evaluate its effectiveness in night or low light environments and the repelling effect on specific birds.

[0203] Ultrasonic bird repelling: Analyze its interference effect on different species of birds and its adaptability in different environments.

[0204] Sound bird repelling: Test the repelling effect of its high - decibel noise or specific frequency sound waves on birds and the possible noise pollution problems.

[0205] B. Analysis of combined strategies:

[0206] Comprehensive effect evaluation: Consider the comprehensive effect when using a combination of multiple bird repelling methods. For example, use laser bird repelling combined with sound bird repelling at night, or use ultrasonic bird repelling combined with visual intimidation means in open areas.

[0207] Resource optimization: Through simulation and experiments, find the optimal combination of bird repelling methods to achieve the best bird repelling effect and resource utilization efficiency.

[0208] C. Risk assessment and response:

[0209] Bird adaptability: Evaluate the risk that birds may develop adaptability to a certain bird repelling method and formulate corresponding countermeasures.

[0210] Equipment failure: Consider the possible failure situations of bird repelling equipment and formulate backup plans to ensure the continuity of bird repelling work.

[0211] The control part is based on the Internet of Things communication module and adopts a decentralized control method. It receives the bird information detected by the detection part, selects a suitable bird repelling strategy, sends a bird repelling instruction to the bird repelling part to achieve the purpose of bird repelling. The Internet of Things platform can connect multiple airport cleaning bird repelling vehicles equipped with intelligent inspection of foreign objects on the airport runway and clearance assistance processing systems, realize information sharing, work together to carry out bird repelling, improve the efficiency and effect of bird repelling. At the same time, the bird repelling data can be recorded and analyzed to discover the laws of birds interfering with aircraft flight, which is convenient for carrying out targeted bird repelling work in the future.

[0212] (5) Cleaning drive module:

[0213] The cleaning drive module is used to remove foreign objects on the runway according to the detection results; the cleaning drive module includes an integrated cleaning device, and the cleaning device is provided with two brush heads 6 rotating in opposite directions, a dry powder sprayer and a sprinkler. It is mainly responsible for cleaning debris such as tire rubber residues, oil stains, and sand and gravel on the runway, while the auxiliary cleaning device is mainly responsible for providing water and dry powder to improve the cleaning efficiency.

[0214] The cleaning device consists of two brush heads 6 arranged front and back, one rotating clockwise and the other counterclockwise. This design brings unique functions and advantages. The auxiliary cleaning device consists of three nozzles. Among them, the middle nozzle is a dry powder sprayer 7, and the other two are sprinklers 8. The working flow chart of the cleaning system is as Figure 4 shown.

[0215] 1) Function of the dry powder sprayer:

[0216] Cleaning of special stains: The dry powder sprayer is suitable for removing special stains such as tire rubber residues, oil stains, and bird droppings that are difficult to clean with water. The sprayed dry powder can effectively adsorb the stains, making the cleaning work more efficient. Selection of dry powder types: Different types of dry powder can be selected according to the type and degree of the stains. For example, for oil stains, dry powder with degreasing effect can be selected; for bird droppings, dry powder with decomposition function can be selected. Control of spraying amount: The dry powder sprayer has an adjustable spraying amount setting, and the appropriate spraying amount can be selected according to the cleaning requirements to avoid waste. Uniform spraying: The sprayer design ensures that the dry powder can be evenly sprayed on the stains to improve the cleaning effect.

[0217] 2) Function of the sprinkler:

[0218] Efficient water spraying and cleaning: The sprinkler can evenly and quickly spray water mist, effectively covering the cleaning area and removing dust, stains, etc. on the ground. Water volume adjustment: The sprinkler has an adjustable water volume setting, and the appropriate water volume can be selected according to different cleaning requirements, such as humidity, stain degree, etc. for cleaning. Automatic navigation cleaning: Combined with the autonomous navigation function, the sprinkler can automatically perform cleaning work according to the preset path, improving the cleaning efficiency. Energy saving and environmental protection: The sprinkler adopts a highly energy-saving design to ensure reducing water resource waste and lowering energy consumption during the cleaning process.

[0219] Both the sprinkler and the dry powder sprayer are integrated with the intelligent control system to achieve one-key switching and intelligent control. Users can remotely control the cleaner through the intelligent operation display panel or the remote control, and adjust the sprinkler or dry powder spray parameters in real time to meet the cleaning needs in different scenarios. According to the type and degree of stains detected by the sensor, it automatically selects to use the sprinkler or dry powder sprayer for cleaning and adjusts the corresponding parameters to achieve the best cleaning effect. The cleaner has waterproof and dustproof functions to ensure normal operation in harsh environments. When the motor or other key components are overloaded, the cleaner will automatically stop working and emit a warning signal to avoid damage. It is equipped with an emergency stop button, and users can immediately stop the cleaner when encountering an emergency.

[0220] Based on the SLAM (Simultaneous Localization and Mapping) algorithm, a fusion solution that mainly uses laser SLAM and supplemented by visual SLAM is adopted and integrated into the control system of the airport cleaning bird repellent vehicle to achieve collaborative work with modules such as the motion control and sensor data acquisition of the airport cleaning bird repellent vehicle. Through the design and optimization of the software architecture, it is ensured that the SLAM algorithm can process sensor data in real time, generate accurate map information, and guide the airport cleaning bird repellent vehicle for precise positioning and navigation.

[0221] The foreign object cleaning path planning adopts an improved RRT* algorithm, considering: the maximum acceleration of the airport cleaning bird repellent vehicle is 2.3 m / s², and the runway occupancy time window (<45 seconds / time).

[0222] (6) Experimental examples:

[0223] Table 7 Experimental data comparison table (based on the airport real dataset ATR-2024) .

[0224] According to the experimental data comparison table shown in Table 7, it can be seen that:

[0225] 1) Significance verification method:

[0226] The paired samples t-test (α = 0.05) is adopted, and the p-value of all indicators <0.01. Calculate Cohen's d effect size:

[0227] The improved d of mAP@0.5 = 1.27 (large effect); the d of energy consumption reduction = 2.03 (extremely large effect).

[0228] 2) The contribution of the spatial attention block (SAB):

[0229] The false alarm rate is reduced: from 15.2 times / hour → 4.8 times / hour; the detection of bird posture adaptability is improved, as shown in Table 8.

[0230] Table 8 Comparison Table of Bird Posture Adaptability 。

[0231] 3) System-level Performance Verification:

[0232] Conducted continuous tests at Chengdu Tianfu International Airport for 30 days:

[0233] Accuracy rate of bird strike early warning: 98.4%; Average response time: 2.3 seconds (industry requirement ≤ 5 seconds); Reduction in the frequency of manual patrols: from 12 times / day → 3 times / day.

[0234] In summary, the key technical points of the present invention are as follows:

[0235] First, the integration part of bird repelling and cleaning:

[0236] 1) Functional integration innovation: Innovatively integrate the bird repelling function and the runway cleaning function into the same platform system. Through a unified control center, realize the coordinated scheduling and management of bird repelling and cleaning operations, break the traditional independent operation mode, and improve the efficiency of airport runway maintenance.

[0237] 2) Intelligent linkage mechanism: Establish an intelligent linkage mechanism for bird repelling and cleaning based on real-time monitoring data. For example, when the platform detects signs of bird activity on the runway, automatically start the bird repelling equipment; at the same time, if it detects foreign objects or pollution that may be brought by bird activity, promptly allocate cleaning resources and quickly carry out cleaning work after bird repelling to ensure runway safety and cleanliness.

[0238] 3) Optimal resource allocation: Dynamically optimize the allocation of bird repelling and cleaning resources based on factors such as the importance of different areas of the runway, the frequency of bird activity, and the probability of foreign object generation. For example, in the key areas of runway takeoff and landing, focus on deploying bird repelling equipment and increasing the cleaning frequency to maximize resource utilization and reduce operating costs.

[0239] Second, the application of image recognition in civil aviation based on YOLOv11:

[0240] 1) Algorithm improvement and optimization: Targeting the characteristics of the civil aviation airport runway environment, make targeted improvements to the YOLOv11 algorithm. For example, optimize the network structure to better adapt to image feature extraction in the complex background of the airport, and improve the recognition accuracy and speed of targets such as runway foreign objects and birds, which includes the adjustment and optimization of key parameters such as the loss function and anchor box settings.

[0241] 2) Multi-source data fusion recognition: Recognize by fusing multiple image data sources. In addition to runway surveillance camera images, it also combines drone inspection images, meteorological monitoring images, etc. Through multi-source data fusion technology, make full use of the advantages of different data sources to improve the recognition robustness of targets such as foreign objects and birds on civil aviation airport runways, especially the recognition ability under complex meteorological conditions.

[0242] 3) Innovation in specific scenario applications: Apply the improved YOLOv11 algorithm to specific scenarios such as foreign object detection and bird monitoring on civil aviation airport runways. By establishing an image dataset for civil aviation scenarios, train and optimize the algorithm so that it can accurately identify various foreign objects (such as metal fragments, plastic garbage, etc.) and different species of birds on the runway, and provide corresponding risk assessment and early warning functions according to the safety requirements of civil aviation.

[0243] 4) Guarantee of real-time performance and stability: In the civil aviation application scenario, ensure that the YOLOv11-based image recognition system has high real-time performance and stability. By adopting hardware acceleration technologies (such as GPU parallel computing) and optimizing the algorithm deployment method, realize the real-time processing of runway images to meet the needs of real-time monitoring of airport runways; at the same time, through redundant design and fault self-diagnosis mechanisms, ensure the stability of the system during long-term operation and avoid potential safety hazards caused by system failures.

Claims

1. An image-based intelligent inspection and clearance assistance processing system for airport runway foreign objects, characterized in that, Including: 1) Image acquisition module: High-definition cameras (1) deployed in the airport runway and clearance area, used to obtain the original image data of the runway and clearance area in real time; 2) Image processing module: An improved YOLOv11 object detection algorithm is used to identify bird, drone or foreign object targets in the clearance area; The architecture of the improved YOLOv11 includes a backbone, a neck component and a head component; The backbone is the main feature extractor, which uses a convolutional neural network to convert the original image data into a multi-scale feature map; the neck component combines feature maps of different scales and transmits them to the head component for prediction; the head component makes object predictions based on the reconstructed feature map; The improvement includes replacing the C2f block in the YOLOv11 neck component with a C3k2 block, and introducing a cross-stage partial and spatial attention block to enhance the small object detection ability; 3) Bird repelling module: Used to repel birds by laser, ultrasonic or sound; 4) Control center: Used to generate a risk assessment report based on the detection results of the image processing module and dispatch the bird repelling module, and the risk assessment report calculates the risk level based on the bird species, density, flight trajectory and foreign object position; 5) Internet of Things communication module: Used to transmit foreign object detection data and bird repelling status to the airport server in real time.

2. The image-based intelligent foreign object inspection and clearance assistance processing system for airport runways according to claim 1, wherein It also includes a cleaning drive module for removing runway foreign objects according to the detection results; the cleaning drive module includes an integrated cleaning device, and the cleaning device is provided with two reversely rotating brush heads (6), a dry powder sprayer (7) and a sprinkler (8).

3. The intelligent inspection of foreign objects on the airport runway and clearance assistance processing system based on images according to claim 2, wherein, It also includes an airport cleaning and bird repelling vehicle (11), the high-definition camera (1) is arranged on the top of the vehicle body, the bird repelling module includes a laser emitter arranged on the pan-tilt (3) on the top of the vehicle body, ultrasonic bird repellers (4) arranged on both sides of the vehicle body and an intelligent induced sound repelling cannon (5); the brush head (6) is arranged at the bottom of the front end of the vehicle body, the dry powder sprayer (7) is arranged in the middle of the front end of the vehicle bottom, and two sprinklers (8) are respectively arranged on both sides of the dry powder sprayer (7); a solar photovoltaic panel (10) is also arranged on the top of the vehicle body; The movement control, positioning and navigation of the airport cleaning and bird repelling vehicle (11) are realized through the control center.

4. The image-based intelligent foreign object detection and clearance assistance processing system for airport runways according to claim 1, characterized in that The C3k2 block includes the following sub-modules: Dynamic snake-shaped convolution: Used to adapt to the geometric structure of the target, by dynamically adjusting the shape of the convolution kernel, enhancing the feature extraction ability for complex foreign object targets in the airport, and dynamically adjusting the shape of the convolution kernel by learning the offset of the input feature map; Cooperative attention convolution: Used to improve the detection accuracy and robustness, by introducing an attention mechanism, enhancing the feature expression of important regions in the feature map, while suppressing unimportant regions, making the network pay more attention to the key parts of the target; Gaussian sampling convolution: Used to sample the convolution kernel through Gaussian distribution, enhancing the smoothness and anti-noise ability of feature extraction; Funnel activation function: Used to provide different gains in different activation intervals, and enhancing the network's adaptability to features of different intensities through a non-linear activation function; Feature fusion adopts a bidirectional feature pyramid strategy, simultaneously utilizing bottom-up and top-down feature information to improve the detection ability for targets of different scales and achieve the fusion and enhancement of multi-scale features.

5. The image-based intelligent foreign object detection and clearance assistance processing system for airport runways according to claim 1, characterized in that The bird repelling module includes a laser bird repelling unit, an ultrasonic bird repelling unit, and a sound bird repelling unit; The laser bird repelling unit emits a green laser beam with a wavelength of 532 nm, and the power range is 500 mW - 5 W. The scanning speed is dynamically adjusted according to the bird density; The ultrasonic bird repelling unit has a working frequency of 12 kHz - 25 kHz and adopts a frequency sweep mode to extend the bird adaptation period; The sound bird repelling unit stores more than 2000 kinds of bird predator calls and screams and supports a random playback strategy.

6. A processing method using the image-based intelligent inspection and clearance assistance processing system for airport runway foreign objects described in claim 2, characterized in that, It includes the following steps: Step 1: Image acquisition; The original image data of the runway and the clear sky area are obtained in real time through high-definition cameras deployed in the airport runway and the clear sky area; Step 2: Improve the main body feature extractor and extract the main body features; The original image data is input into the main body feature extractor, and the original image data is converted into a multi-scale feature map through a convolutional neural network; the main body feature extractor is the backbone of the improved YOLOv11 architecture; the improvement includes: replacing the standard convolutions in the 3rd, 7th, and 11th layers of YOLOv11 with dynamic snake-shaped convolutions, and embedding a collaborative attention mechanism in the C3 module; Step 3: Reconstruct the neck component and complete the intermediate processing; First, replace the C2f block in the neck component of the YOLOv11 architecture with a C3k2 block. Use multiple C3k2 blocks to sample and connect feature maps of different scales, aggregate and enhance the feature representations at different scales, and complete the reconstruction of the feature map; Step 4: Improve the head component and complete object localization and output; Introduce a new cross-stage part and a spatial attention block in the head component of the YOLOv11 architecture; based on the original spatial pyramid pooling-fast block of the YOLOv11 architecture, as well as the newly introduced cross-stage part and spatial attention block, obtain the detection results of birds, drones, or foreign objects in the clear sky area based on the feature map reconstructed by the neck component; Step 5: Establish a risk assessment model, calculate the threat level according to the detection results through the constructed threat level function, mobilize the bird repelling module to select the corresponding bird repelling strategy for bird repelling according to the threat level, or mobilize the cleaning drive module to clean foreign objects.

7. The processing method of the image-based intelligent foreign object detection and clearance assistance processing system for airport runways according to claim 6, characterized in that, In Step 2, converting the original image data into a multi-scale feature map through a convolutional neural network includes using the convolutional neural network for path prediction. The required data types include: Path trajectory data: including points arranged in chronological order during the cleaning process, and each point represents a longitude and latitude coordinate, speed, and direction information on the path; Image representation: Convert the cleaning work path data into an image form, simplify the three-dimensional space into a two-dimensional plane, draw the path on a two-dimensional grid, and then use it as the input of the convolutional neural network; Vector sequence: Convert the path trajectory data into a vector sequence, and each vector contains the position information of a certain point on the path and other related features; Environmental map data: It is used to provide the context information required for path prediction, help the convolutional neural network understand the relationship between the path and the environment, and also encrypt the data.

8. The processing method of the image-based intelligent inspection and clearance assistance processing system for airport runway foreign objects according to claim 6, characterized in that, In step 4, the spatial attention block includes the following sub-modules: position-sensitive convolution and channel weight generator; the position-sensitive convolution is used to introduce position information in the convolution operation to enhance the perception ability of the target spatial position; the channel weight generator is used to compress the number of channels of the input feature map through convolution according to the spatial weight; Spatial weight matrix W ij The calculation formula is as follows: ； Among them, σ L is the LeakyReLU activation function, is the value of the intermediate feature at position (i, j), H is the height of the feature map, W is the width of the feature map, is the value of the intermediate feature at position (k, l).

9. The processing method of the image-based intelligent inspection and clearance assistance processing system for airport runway foreign objects according to claim 6, characterized in that, In step 4, object localization is used to locate the bounding box coordinates of the object in the image, specifically as follows: Step 4.1: Configure the detection head architecture; Set 3 cascaded CBS blocks for each detection branch, respectively performing feature-preserving convolution, channel compression, and spatial feature enhancement. Insert cross-stage residual connections after each layer, and batch normalization implements dynamic momentum batch normalization, and the momentum coefficient changes with the number of training epochs; Step 4.2: Coordinate regression; Map the feature map to the anchor box: Divide the grid of the feature map reconstructed by the input neck component, and each grid point is associated with 3 prior anchor boxes; Calculate the convolution output dimension: Calculate the number of output channels of each grid point and perform non-linear decoding on the bounding box coordinates: ； ； ； ； Among them, t x , t y , t w and t h are the original outputs of the network after being activated by the Silu activation function; c x and c y are the coordinates of the grid point at the upper left corner of the feature map; w a and h a are the reference sizes of the anchor boxes; is the Sigmoid activation function; b x is the abscissa of the center point of the bounding box, b y is the ordinate of the center point of the bounding box, b w is the width of the bounding box, b h is the height of the bounding box; The dynamic scaling factor introduces a learnable scale coefficient λ and is updated through gradient descent: ； Among them, is the updated scale coefficient, is the learnable scale coefficient; is the learning rate; is the loss gradient.

10. The processing method of the image-based intelligent foreign object detection and airspace assistance processing system for airport runways according to claim 6, characterized in that, In step 5, a risk assessment model is established, specifically including: Step 5.1: Construct a dynamic target feature library; Obtain the detection results of the improved YOLOv11, including the category, location, size, and speed of the detected objects, and construct feature vectors according to the time series : ； where x and y are the horizontal and vertical position coordinates of the target on the image plane, w and h are the width and height of the detected target, v x and v y are the horizontal and vertical velocity components of the dynamically monitored target on the image plane, class is the category to which the target belongs, and confidence represents the confidence level of the model in the detection result, which is a value between 0 and 1. The higher the value, the more certain the model is that the detected target belongs to the specified category; Step 5.2: The threat degree function is constructed as: ； Among them, T is the threat level of the target, which is used to evaluate the potential threat degree of the target to the safety of the airport runway, S size , V speed , D distance and E env are respectively the volume, speed, distance between the target and the key area of the runway, and the environmental factor value, , , and are respectively the corresponding weight coefficients.

Images