Airport scene multi-scale target detection method based on enhanced target detection model

Abstract

The invention provides an airport scene multi-scale target detection method based on an enhanced target detection model, and relates to the technical field of target detection, and the method comprises the steps: building an airport scene multi-scale moving target data set; constructing an enhanced target detection model by using a hierarchical shift window self-attention mechanism; training the enhanced target detection model by using the airport scene multi-scale moving target data set to obtain a trained enhanced target detection model; and detecting the airport scene target data by using the trained enhanced target detection model to obtain a target detection result, and completing the detection of the airport scene multi-scale target. According to the method, the problem that multi-scale target detection is difficult to accurately carry out for an airport complex scene is solved.

Description

Technical Field

[0001] This specification relates to the field of target detection technology, and in particular to a multi-scale target detection method for airport scenes based on an enhanced target detection model. Background Technology

[0002] With the rapid expansion of global air transport networks, the scale of airport infrastructure and the frequency of aircraft takeoffs and landings have continued to rise, leading to a simultaneous increase in the density of ground service vehicles, the number of resident aircraft, and the intensity of ground staff movement. This complex interaction of multimodal moving targets significantly increases the operational risks at airports, typically manifesting as aircraft taxiing conflicts, aircraft-vehicle encounter risks, and ground staff intrusion into critical areas. Against this backdrop, computer vision-based moving target detection and real-time positioning technology has become the core of airport intelligent surveillance systems, with its detection accuracy and reliability directly determining the safety margin and efficiency of airport operations.

[0003] Airport target detection currently faces three major technical bottlenecks. First, there's the problem of extreme scale variation: airport scenes simultaneously contain large-scale aircraft (up to 70 meters wide) and small-scale critical safety targets (such as ground crew members 1 meter tall), often coexisting in the same image frame. The feature abstraction process of deep convolutional neural networks is accompanied by spatial resolution decay, causing key information about small targets to be lost in the high-level semantic space. Second, there's the imbalance in the long-tail distribution of categories: target categories exhibit an unbalanced distribution. The difference in sample size between aircraft (high-frequency category) and ground crew / special vehicles (low-frequency category) leads to the model optimization process being dominated by the dominant category, significantly reducing the recall rate of low-frequency targets. Finally, there are complex environmental interferences: airport background noise (such as runway markings and lighting systems) and target occlusion effects further exacerbate the difficulty of feature extraction for small-scale targets.

[0004] Existing technologies also have the following drawbacks: 1. Loss of small target information: Gridded prediction and progressive downsampling cause spatial details of small targets (such as ground staff) to be lost in deep feature maps. 2. Static scale adaptation: The hierarchical division of FPN / PAN is fixed and cannot dynamically adapt to the problem of coexistence of extreme-scale targets in airport scenarios. 3. Sensitivity to class imbalance: Standard label allocation strategies ignore long-tail distribution, resulting in low recall rates for low-frequency categories (personnel). Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, the multi-scale target detection method for airport scenes based on an enhanced target detection model provided by this invention solves the problem of accurately performing multi-scale target detection in complex airport scenes.

[0006] To achieve the aforementioned objectives, the present invention employs the following technical solution: a multi-scale target detection method for airport scenes based on an enhanced target detection model, comprising: S1: Establish a multi-scale moving target dataset for airport surfaces; S2: Construct an enhanced object detection model using a hierarchical shift window self-attention mechanism; S3: Using the multi-scale moving target dataset of the airport scene, train the enhanced target detection model to obtain the trained enhanced target detection model; S4: Use the trained enhanced target detection model to detect targets in airport scene data, obtain target detection results, and complete the detection of multi-scale targets in airport scene.

[0007] Furthermore, the enhanced target detection model includes: The backbone network layer is used to perform self-attention calculation on the multi-scale moving target dataset of the airport scene through the hierarchical shift window self-attention mechanism, retaining shallow high-resolution features and enhancing the ability to extract fine-grained features of small targets, thus obtaining enhanced features. The detection head layer includes an adaptive receptive field module, which uses deformable convolutions in the adaptive receptive field module to dynamically adjust the fusion range of the enhanced features to obtain fused features; The head prediction layer is used to perform object detection on the fused features at each location to obtain the object detection result.

[0008] Furthermore, the expression for the fusion range of the enhanced features is: ; in, This represents the output feature map of the k-th layer after processing by the Adaptive Receptive Field (ARF) module. This represents a deformable convolution with a learned offset. Represents the channel attention weight vector. This indicates element-wise multiplication. Let i represent the input feature map at the i-th scale. This indicates that it has a learning offset; The expression for the centrality score of the target detection result is: ; in, This represents the predicted centrality score at spatial location (x, y). This represents the function that takes the minimum value. This represents the function that takes the maximum value. This represents the lateral distance from the predicted point (x, y) to the left boundary of the predicted bounding box. This represents the lateral distance from the predicted point (x, y) to the right boundary of the predicted bounding box. This represents the vertical distance from the predicted point (x, y) to the upper boundary of the predicted bounding box. This represents the vertical distance from the predicted point (x, y) to the lower boundary of the predicted bounding box.

[0009] Further, S3 includes: Using a multi-scale moving target dataset from an airport surface, small target enhancement was performed on the enhanced target detection model to obtain small target enhancement results; Based on the small target augmentation results, the size is adjusted using bilinear interpolation, and the sampling probability of small targets is dynamically increased using scale-aware mosaic sampling to obtain the augmented dataset; Based on the augmented dataset, a multi-target training method is used to obtain a well-trained augmented target detection model.

[0010] Furthermore, the sample allocation mechanism includes: Based on the augmented dataset, a cost matrix is ​​constructed using the optimal transmission allocation strategy; By solving the cost matrix, the optimal matching result is obtained. The enhanced target detection model is then trained on multiple targets to obtain a well-trained enhanced target detection model.

[0011] Furthermore, the expression for the cost matrix is: ; in, This represents the value in the i-th row and j-th column of the cost matrix. The weights represent the classification loss. Indicates focal loss. This represents the probability distribution of the class prediction for the i-th prediction box. This represents the true category label of the j-th true bounding box. The weights representing the regression loss. Indicates CIoU loss. This represents the bounding box coordinates of the i-th predicted box by the model. Represents the bounding box coordinates of the j-th ground truth bounding box; The expression for the optimal matching result is: ; in, This represents the optimal matching result. This indicates that the constraints are met. All transmission plans P Among them, find the one that minimizes the objective function. P , Indicates the transportation plan, Representing the cost matrix, Representation matrix P and C The inner product, This represents the entropy regularization term.

[0012] Furthermore, the small target enhancement process includes: Extracting the area of ​​target objects from a multi-scale moving target dataset at airport surfaces; For each target object area, generate multiple duplicate samples; Apply an affine transformation to the duplicate samples and paste them to obtain the corresponding new coordinates; Based on the new coordinates of each repeated sample, image fusion is performed on the repeated samples to obtain the small target enhancement result.

[0013] Furthermore, the expression for the new coordinates is: ; in, Indicates the new center point coordinates. Represents the original x-coordinate. This represents the horizontal coordinate offset randomly generated within the image range. Represents the original position's ordinate. This represents the offset of the ordinate position randomly generated within the image range; The expression for the small target enhancement result is: ; in, This indicates that smaller goals enhance the results. Indicates the transparency blending factor. Represents the foreground image. This represents the background image.

[0014] Furthermore, the expression for the augmented dataset is: ; ; ; ; ; in, This represents the input scale randomly selected in the t-th training iteration. This indicates rounding to the nearest integer. This indicates the base resolution, which defaults to 640 pixels. This represents the scale factor, set to 0.5. This represents a random number sampled from a uniform distribution U(0,1). Indicates uniform distribution. This represents the pixel value at position (x', y') of the resized image. and This represents the coordinates in the image after resizing. Indicates the position of the original image ( , The pixel value at () This represents the function that takes the maximum value. This indicates the adjusted image width. This indicates the adjusted image height. Indicates the scale of dynamic selection. Indicates the height of the original image. Indicates the width of the original image. This represents the spatial attention weight at position (i,j). This represents the sigmoid activation function. This represents the total number of channels in the feature map, and c represents the loop variable used to iterate through the channel indices. This represents the value of the feature map at position (i,j) and channel c. This indicates the probability that an image containing small objects was sampled. This indicates the number of images in the dataset that contain small objects. This represents the total number of images in the dataset.

[0015] The beneficial effects of this invention are as follows: This invention provides a multi-scale target detection method for airport scenes based on an enhanced target detection model. By constructing an enhanced target detection model, robustness enhancement is achieved in multi-scale target collaborative detection. Relying on the sparse computation characteristics and hierarchical design of deformable convolution in the enhanced target detection model, real-time lightweighting is achieved while maintaining high accuracy. By training the enhanced target detection model, for long-tailed distribution of categories, density-aware sample allocation and weighted loss are used to effectively alleviate the bottleneck of low-frequency category detection. The generalization ability of the enhanced target detection model is significantly enhanced, and the detection integrity in complex scenes is verified through visualization. Attached Figure Description

[0016] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is an exemplary flowchart of a multi-scale target detection method for airport scenes based on an enhanced target detection model, as shown in some embodiments of this specification. Figure 2 This is an exemplary schematic diagram of boxlines for logarithmically scaled pixel areas of “people,” “vehicles,” and “aircraft” (ASMOD dataset) according to some embodiments of this specification; Figure 3This is an exemplary schematic diagram of the detailed architecture of Aero YOLO according to some embodiments of this specification; Figure 4 This is an exemplary schematic diagram showing a comparison of detection results on the NWPU-VHR10 dataset according to some embodiments of this specification; Figure 5 This is an exemplary schematic diagram of the precision-recall curves of Aero YOLO and competitive detectors on the ASMOD dataset, as shown in some embodiments of this specification. Figure 6 This is an exemplary schematic diagram showing a comparison of detection results on the AMOD dataset according to some embodiments of this specification. Detailed Implementation

[0017] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0018] Example Figure 1 This is an exemplary flowchart illustrating a multi-scale target detection method for airport scenes based on an enhanced target detection model, according to some embodiments of this specification. Figure 1 As shown, the process includes the following steps. In some embodiments, the process may be executed by a processor.

[0019] S1: Establish a multi-scale moving target dataset for airport surfaces.

[0020] The Airport Scene Multi-Scale Moving Target Dataset is a multi-scale image dataset used to train augmented target detection models. For example, the Airport Scene Multi-Scale Moving Target Dataset may include 2,429 high-resolution images, labeled with 3,429 instances (1,240 aircraft, 1,455 special vehicles, and 734 ground crew members), and divided into training, validation, and test sets in a 7:1:2 ratio, as shown in Table 1. Scale coverage includes continuous distribution of target pixel areas (e.g., ...). Figure 2 As shown), personnel (10³ to 10 4 Pixels), vehicles (10) 4 Up to 10 5 Pixels), aircraft (10) 6 (pixels), simulating extreme scale coexistence scenarios.

[0021] In some embodiments, the annotation system for the airport surface multi-scale moving target dataset includes: fine-grained bounding box annotations for three types of targets: aircraft, special vehicles, and ground staff, with additional descriptions of spatiotemporal related events (such as aircraft-vehicle interactions).

[0022] Table 1 Instance Count of Each Category

[0023] S2: Construct an enhanced object detection model using a hierarchical shift window self-attention mechanism.

[0024] Augmented target detection models are neural network models used for detecting targets on airport surfaces. For example, an augmented target detection model can be a neural network model based on the YOLOv11 architecture with triple collaborative improvements.

[0025] In some embodiments, the enhanced target detection model includes a backbone network layer, a detection head layer, and a head prediction layer.

[0026] The backbone network layer is used to perform self-attention calculation on the multi-scale moving target dataset of the airport scene through a hierarchical shift window self-attention mechanism, preserving shallow high-resolution features and enhancing the ability to extract fine-grained features of small targets, thus obtaining enhanced features.

[0027] Enhanced features are those that retain shallow high-resolution features and enhance the ability to extract fine-grained features from small targets (such as ground crew).

[0028] In some embodiments, the backbone network layer can be replaced by a Swing Transformer to replace the traditional convolutional neural network (CNN), which preserves shallow high-resolution features through a hierarchical shift window self-attention mechanism, thereby enhancing the ability to extract fine-grained features of small targets (such as ground crew).

[0029] In some embodiments, the expression for computation by the self-attention mechanism can be: ; in, This indicates the self-attention mechanism. Represents the query matrix. Represents the key matrix. Represents a value matrix, express function, Represents the dimension of a vector. To represent the transpose of a matrix, This indicates a relative position offset.

[0030] In airport target detection, Q can be understood as "the location of current interest", K as "other potentially relevant locations", and V as "the actual features of these locations".

[0031] In some embodiments, Indicates relative position offset, via Cross-window interaction can be achieved by cyclically shifting within a local window.

[0032] In some embodiments, the Pyramid Visual Transformer (PVT) or Twins Transformer can be considered as an alternative to the Swing Transformer in the backbone network design. PVT preserves multi-scale features through a progressive feature pyramid, avoiding the computational overhead of window shifting operations; Twins Transformer integrates a local-global attention mechanism, reducing computational complexity while maintaining high-resolution feature representation. Both can handle extreme-scale targets more flexibly and alleviate the problem of boundary information loss. For lightweight deployment scenarios, ConvNeXt or EfficientNetV2 architectures can be selected. ConvNeXt achieves a Transformer-like transformation of CNNs through depthwise separable convolutions, significantly compressing the number of parameters; EfficientNetV2 optimizes the accuracy-speed balance based on a composite scaling strategy, adapting to the real-time requirements of airport edge computing devices.

[0033] The detection head layer includes an adaptive receptive field module, which uses deformable convolutions in the adaptive receptive field module to dynamically adjust the fusion range of the enhanced features to obtain fused features.

[0034] Fusion features are features obtained by fusing multiple enhancement features.

[0035] In some embodiments, the detection head layer includes an adaptive receptive field module that dynamically adjusts the feature fusion range using deformable convolution to obtain fused features.

[0036] In some embodiments, the expression for the fusion range of the enhanced features can be: ; in, This represents the output feature map of the k-th layer after processing by the Adaptive Receptive Field (ARF) module. This represents a deformable convolution with a learned offset. Represents the channel attention weight vector. This indicates element-wise multiplication. Let i represent the input feature map at the i-th scale. This indicates that there is a learning offset.

[0037] The head prediction layer is used to perform object detection on the fused features at each location to obtain the object detection result.

[0038] The target detection results are the results of detecting targets on the airport surface. For example, the target detection results may include classification scores. Bounding box offset Centrality score .

[0039] In some embodiments, the expression for the centrality score of the object detection result can be: ; in, This represents the predicted centrality score at spatial location (x, y). This represents the function that takes the minimum value. This represents the function that takes the maximum value. This represents the lateral distance from the predicted point (x, y) to the left boundary of the predicted bounding box. This represents the lateral distance from the predicted point (x, y) to the right boundary of the predicted bounding box. This represents the vertical distance from the predicted point (x, y) to the upper boundary of the predicted bounding box. This represents the vertical distance from the predicted point (x, y) to the lower boundary of the predicted bounding box.

[0040] In some embodiments, dynamic convolution can replace deformable convolution in the optimization of the adaptive receptive field module. By dynamically adjusting parameters through the fusion of weights from multiple convolutional kernels, offset prediction errors can be avoided, thereby improving the positioning accuracy of large targets (such as aircraft). Attention-driven schemes can employ spatial attention modules (SAM) or hybrid domain attention mechanisms (CBAM). The former strengthens the features of small target regions through spatial weight maps, while the latter combines spatial and channel dimensions to enhance feature discriminativeness; both can effectively suppress background interference.

[0041] S3: Using the multi-scale moving target dataset of the airport scene, train the enhanced target detection model to obtain the trained enhanced target detection model.

[0042] In some embodiments, the processor can use a multi-scale moving target dataset of an airport scene to perform small target enhancement on the enhanced target detection model to obtain small target enhancement results; based on the small target enhancement results, the size is adjusted using bilinear interpolation, and the sampling probability of small targets is dynamically increased using scale-aware mosaic sampling to obtain an enhanced dataset; based on the enhanced dataset, multi-target training is performed through a sample allocation mechanism to obtain a trained enhanced target detection model.

[0043] Small target enhancement results are the results of enhancing the area image of small target objects.

[0044] Augmented datasets are datasets that have been resized and have increased sampling probability for small targets.

[0045] In some embodiments, the sample allocation mechanism process may include: constructing a cost matrix based on the augmented dataset and using an optimal transfer allocation strategy; obtaining the optimal matching result by solving the cost matrix; and performing multi-target training on the augmented target detection model to obtain a trained augmented target detection model.

[0046] The cost matrix is ​​a matrix used to determine how tags are assigned.

[0047] In some embodiments, the cost matrix can be expressed as: ; in, This represents the value in the i-th row and j-th column of the cost matrix. The weights represent the classification loss. Indicates focal loss. This represents the probability distribution of the class prediction for the i-th prediction box. This represents the true category label of the j-th true bounding box. The weights representing the regression loss. Indicates CIoU loss. This represents the bounding box coordinates of the i-th predicted box by the model. This represents the bounding box coordinates of the j-th ground truth bounding box.

[0048] The optimal matching result is the best matching result among the sample assignments. For example, the optimal matching result... P It is an optimal allocation matrix, whose elements Pij Indicates the first i The prediction box is dynamically assigned to the first... j The association strength or probability of each ground truth bounding box. In actual training, for each ground truth object, the OTA policy will assign it a previous association strength or probability. k (For example k =10) of the lowest-cost predicted boxes are used as positive samples. This allocation method is not a static one-to-one matching, but a dynamic one-to-many matching, and it will be dynamically adjusted according to the frequency of the category. For categories that appear less frequently in the dataset (such as ground staff), the system will tend to allocate more positive samples to them, thereby effectively alleviating the class imbalance problem and significantly improving the recall rate of low-frequency categories.

[0049] In some embodiments, the processor may be provided with A real label and The prediction results are used to solve for the optimal match using the Sinkhorn-Knopp algorithm. For each true label, OTA is selected before... For each matching item, the frequency of the matching is adjusted by the reciprocal of the frequency of the category to obtain the optimal matching result.

[0050] In some embodiments, the expression for the optimal matching result can be: ; in, This represents the optimal matching result. This indicates that the constraints are met. Among all transmission plans P, find the one P that minimizes the objective function. Indicates the transportation plan, Representing the cost matrix, This represents the inner product of matrices P and C. This represents the entropy regularization term. It is the total cost under a certain allocation scheme PP, and the optimization objective is to minimize this total cost. It is the information entropy of the transmission plan PP. The larger the entropy value, the more "smooth" or "dispersed" the allocation scheme PP is; the smaller the entropy value, the more "sharp" or "concentrated" the allocation scheme is.

[0051] In some embodiments, the loss function for training the enhanced object detection model can be the cross-entropy loss function.

[0052] In some embodiments, for the label allocation strategy, the optimal transport allocation (OTA) can be optimized into the SimOTA scheme, which uses Top-k cost matching to replace the Sinkhorn algorithm to reduce computational complexity and significantly improve the training efficiency of long-tailed distributed data (such as ground staff samples). Adaptive training sample selection (ATSS) dynamically adjusts the positive and negative sample thresholds based on the target statistical characteristics, avoiding the limitations of preset anchor boxes and achieving adaptive mitigation of class imbalance. At the loss function level, a gradient harmonic mechanism (GHM) or an improved Focal Loss can be introduced. The former focuses on difficult samples (such as occluded vehicles) by suppressing the gradient of easily classified samples, while the latter dynamically adjusts the α coefficient to balance class weights.

[0053] In some embodiments, the small target enhancement process may include: extracting the area of ​​target objects from a multi-scale moving target dataset of an airport scene; generating multiple duplicate samples for each target object area; applying an affine transformation to the duplicate samples and pasting them to obtain corresponding new coordinates; and performing image fusion on the duplicate samples based on the new coordinates of each duplicate sample to obtain the small target enhancement result.

[0054] In some embodiments, the expression for the new coordinates can be: ; in, Indicates the new center point coordinates. Represents the original x-coordinate. This represents the horizontal coordinate offset randomly generated within the image range. Represents the original position's ordinate. This represents the offset of the ordinate position randomly generated within the image range.

[0055] In some embodiments, the expression for the small target enhancement result can be: ; in, This indicates that smaller goals enhance the results. Indicates the transparency blending factor. Represents the foreground image. This represents the background image.

[0056] In some embodiments, the processor can be based on a multi-scale moving target dataset of an airport surface, with a resolution of [resolution value missing] for each batch. from Medium-resolution sampling, resizing via bilinear interpolation while maintaining aspect ratio, and scale-aware mosaic sampling dynamically increase the sampling probability of small targets to obtain an augmented dataset; among which, Indicates the reference resolution.

[0057] In some embodiments, the baseline probability is set to 0.5 to ensure that sampling is still enforced even when the proportion of small targets is extremely low, avoiding missed detections. A scaling factor of 0.3 is used to dynamically adjust the probability weights based on the actual proportion of small targets in the ASMOD dataset (approximately 30%), thereby balancing the need for sample diversity and class balance, and optimizing the model's focus on small targets. This ensures that low-proportion targets fully participate in training.

[0058] In some embodiments, the Dynamic Scale Augmentation (DSA) of the data augmentation module can be extended to a multi-scale hybrid augmentation technique. This technique generates synthetic samples by fusing images of different scales, simulating the scale continuity of targets in real-world scenes (such as the coexistence of aircraft and vehicles), and enhancing the model's robustness to scale abrupt changes. To address the scarcity of small target samples, StyleGAN can be used to generate high-fidelity small target images and fuse them into the real background, or SRGAN super-resolution technology can be introduced to reconstruct low-resolution regions, enhancing the detailed features of small targets while avoiding contextual loss due to local magnification.

[0059] In some embodiments, the expression for augmenting the dataset can be: ; ; ; ; ; in, This represents the input scale randomly selected in the t-th training iteration. This indicates rounding to the nearest integer. This indicates the base resolution, which defaults to 640 pixels. This represents the scale factor, set to 0.5. This represents a random number sampled from a uniform distribution U(0,1). Indicates uniform distribution. This represents the pixel value at position (x', y') of the resized image. and This represents the coordinates in the image after resizing. Indicates the position of the original image ( , The pixel value at () This represents the function that takes the maximum value. This indicates the adjusted image width. This indicates the adjusted image height. Indicates the scale of dynamic selection. Indicates the height of the original image. Indicates the width of the original image. This represents the spatial attention weight at position (i,j). This represents the sigmoid activation function. This represents the total number of channels in the feature map, and c represents the loop variable used to iterate through the channel indices (from 1 to C). This represents the value of the feature map at position (i,j) and channel c. This indicates the probability that an image containing small objects was sampled. This indicates the number of images in the dataset that contain small objects. This represents the total number of images in the dataset.

[0060] In some embodiments, at the overall architecture level, anchor-free detection frameworks such as FCOS avoid scale bias of preset anchor boxes through center point prediction mechanisms, naturally adapting to the scale differences between aircraft and personnel. Query-based detectors such as Sparse R-CNN, combined with learnable queries and iterative refinement mechanisms, or improved DETR equipped with multi-scale deformable attention, can effectively improve the recall rate of small targets and optimize the detection efficiency of dense targets (such as vehicles on the tarmac). Multi-stage solutions can construct a joint optimization framework of Cascade R-CNN and multi-scale ROIs, significantly improving localization accuracy through cascaded bounding box refinement and hierarchical feature pooling, meeting the needs of safety-critical scenarios such as runway intrusion detection.

[0061] By maintaining high-resolution features through the Swin Transformer backbone network and employing shifted window self-attention to enhance small target detail capture, combined with the deformable convolution of the ARF module to dynamically focus on the target region, the detection accuracy of small targets is significantly improved (AP@0.5 for ground crew reaches 87.69%, a 3.44% improvement over the benchmark). The innovative fusion of adaptive receptive field adjustment from the ARF module and joint loss optimization using the OTA strategy achieves robust enhancement in multi-scale target collaborative detection (mAP@0.5 reaches 93.59%, with simultaneous optimization of extreme-scale targets such as aircraft AP to 99.75% and personnel AP to 87.69%). Leveraging the sparse computational characteristics of deformable convolution and the hierarchical design of the Swin Transformer, real-time lightweight operation is achieved while maintaining high accuracy (mAP improved by 4.0% at the same FPS) (84.33 FPS, 7.0M parameters, 27.3M computation). For long-tailed category distributions, the OTA strategy effectively alleviates the bottleneck of low-frequency category detection through density-aware sample allocation and weighted loss (personnel AP improved by 3.44%). Multi-scale augmentation training based on the ASMOD continuous-scale dataset and the DSA strategy significantly enhances the model's generalization ability (achieving mAP@0.5 98.38% in cross-domain testing on NWPU VHR-10), and the completeness of detection in complex scenarios is verified through visualization. The results have been rigorously verified through controlled experiments, ablation analysis, and cross-domain testing, establishing it as a high-performance solution for the aviation inspection field.

[0062] S4: Use the trained enhanced target detection model to detect targets in airport scene data, obtain target detection results, and complete the detection of multi-scale targets in airport scene.

[0063] In some embodiments, the processor can use a trained enhanced target detection model to detect airport surface target data. The backbone network layer uses a hierarchical shift window self-attention mechanism to perform self-attention calculations on the multi-scale moving target dataset of the airport surface, preserving shallow high-resolution features and enhancing the fine-grained feature extraction capability for small targets to obtain enhanced features. The detection head layer, including an adaptive receptive field module, uses deformable convolutions within the adaptive receptive field module to dynamically adjust the fusion range of the enhanced features to obtain fused features. Finally, the head prediction layer performs target detection on the fused features at each location to obtain the target detection result.

[0064] In some embodiments, such as Figure 4 , Figure 5 and Figure 6As shown, the processor can use validation and test sets to perform performance verification on the trained augmented object detection model, obtaining the performance verification results. For example, the accuracy metric: 93.59% mAP@0.5 on the ASMOD dataset, an improvement of 1.88% compared to YOLOv11 (Table 2); the efficiency metric: AP for ground staff detection improved by 3.44%. The dataset contains 7.0M datasets, requires 27.3 GFLOPs of computation, and has an inference speed of 84.33 FPS (Table 2), meeting the requirements for real-time monitoring. Generalization ability: 98.38% mAP@0.5 on the NWPU VHR-10 remote sensing dataset (Table 3), validating cross-domain adaptability.

[0065] Table 2 shows the comparison with state-of-the-art methods on the ASMOD dataset.

[0066] Table 3 shows the performance comparison on the NWPU VHR-10 dataset.

[0067] The Aero YOLO framework implements a triple collaborative optimization based on the YOLOv11 architecture: (1) Constructing a hierarchical SwinTransformer backbone network, achieving multi-scale feature pyramid fusion through a shift window attention mechanism; (2) Designing an Adaptive Receptive Field (ARF) module as the core component of the neck network, dynamically adjusting the feature fusion range using deformable convolution; (3) Introducing an Optimal TransportAssignment (OTA) strategy to establish a density-aware sample allocation mechanism to cope with long-tailed class distributions. Figure 3 As shown, this framework integrates a Dynamic Scale Augmentation (DSA) strategy, enhancing the model's robustness to scale variations through multi-scale training and enhanced mosaic techniques. The Swin Transformer is used as the core feature extractor. Its hierarchical design (four stages) and shift-window attention mechanism effectively capture global contextual information at different semantic levels, particularly beneficial for preserving features of small targets. An innovative Adaptive Receptive Field (ARF) module is introduced for multi-scale feature fusion. This module integrates deformable convolution and scale-aware attention mechanisms to dynamically adjust the receptive field, adaptively handling extreme-scale targets ranging from large aircraft to small pedestrians. Finally, the detection component outputs the target's class score and bounding box coordinates to complete the recognition and localization tasks. Through dynamic adaptive optimization across the three stages of feature extraction, fusion, and allocation, the system addresses the complex challenges of extreme scale variation and class imbalance in airport surface surveillance.

[0068] To address the issue of lost spatial details of small targets due to gridded prediction and progressive downsampling in the YOLO series architecture, a hierarchical attention-enhanced feature extraction mechanism is constructed. A Swin Transformer backbone network is used to preserve high-resolution features in shallow layers, and a shift-window self-attention mechanism is employed to strengthen the perception of local details of small targets, thereby eliminating information annihilation in deep semantic spaces. To address the adaptive limitations of fixed-stride anchor boxes and static feature pyramids (such as FPN / PAN) in scenarios where extreme-scale targets coexist, a deformable convolution-driven dynamic receptive field adapter is designed. By learning offsets, the sampling position of the convolution kernel is dynamically adjusted, achieving coordinated adaptation between the feature extraction range and the target scale, significantly improving the robustness of multi-scale representation in complex backgrounds.

[0069] To address the computational efficiency bottleneck and small target omission issues caused by the global self-attention mechanism in Transformer architectures such as DETR, this paper optimizes the network computation path by embedding a lightweight deformable convolutional module in the detection head. Inference speed is improved through local sparse attention computation, and a cross-scale selective fusion mechanism is combined to enhance the ability to capture sub-pixel-level targets. To address the class mismatch and spatial relationship deficiencies in general datasets (such as KITTI and UTD-19), an airport-specific benchmark dataset ASMOD is constructed. This dataset covers fine-grained bounding box annotations for three key target categories: aircraft, special vehicles, and ground staff. Its continuous scale distribution characteristics (target pixel area spanning 10³-10⁻¹⁰) are also addressed. 6 The annotation of events (at the magnitude of the event) and spatiotemporal related events provides domain-adaptive supervision signals for model training.

[0070] To address the sample imbalance problem of low-frequency categories (such as ground staff) caused by the long-tail distribution of airport targets, a dynamic label allocation strategy based on optimal transport theory is introduced. The Sinkhorn-Knopp algorithm is used to jointly optimize classification and regression losses, and positive sample selection is dynamically weighted based on category frequency, effectively mitigating the dominant effect of the dominant category on model optimization. To address the limitations of traditional data augmentation methods in cross-scale scenarios, a dynamic scale augmentation training strategy is proposed, integrating multi-resolution random sampling, cross-context instance transfer, and scale-aware mosaic sampling mechanisms. By constructing a data distribution simulation environment covering continuous scale space, the model's generalization ability to scale variations and background interference is enhanced.

[0071] In some embodiments of this specification, a multi-scale target detection method for airport scenes based on an enhanced target detection model is provided. By constructing an enhanced target detection model, robustness enhancement is achieved in multi-scale collaborative target detection. Relying on the sparse computational characteristics and hierarchical design of deformable convolution in the enhanced target detection model, real-time lightweighting is achieved while maintaining high accuracy. By training the enhanced target detection model, density-aware sample allocation and weighted loss are used to effectively alleviate the bottleneck of low-frequency category detection for long-tailed category distribution. The generalization ability of the enhanced target detection model is significantly enhanced, and the detection integrity in complex scenes is verified through visualization.

Claims

1. A multi-scale target detection method for airport scenes based on an enhanced target detection model, characterized in that, include: S1: Establish a multi-scale moving target dataset for airport surfaces; S2: Construct an enhanced object detection model using a hierarchical shift window self-attention mechanism; S3: Using the multi-scale moving target dataset of the airport scene, train the enhanced target detection model to obtain the trained enhanced target detection model; S4: Use the trained enhanced target detection model to detect targets in airport scene data, obtain target detection results, and complete the detection of multi-scale targets in airport scene.

2. The multi-scale target detection method for airport scenes based on an enhanced target detection model according to claim 1, characterized in that, The enhanced target detection model includes: The backbone network layer is used to perform self-attention calculation on the multi-scale moving target dataset of the airport scene through the hierarchical shift window self-attention mechanism, retaining shallow high-resolution features and enhancing the ability to extract fine-grained features of small targets, thus obtaining enhanced features. The detection head layer includes an adaptive receptive field module, which uses deformable convolutions in the adaptive receptive field module to dynamically adjust the fusion range of the enhanced features to obtain fused features; The head prediction layer is used to perform object detection on the fused features at each location to obtain the object detection result.

3. The multi-scale target detection method for airport scenes based on an enhanced target detection model according to claim 2, characterized in that, The expression for the fusion range of the enhanced features is: ； in, This represents the output feature map of the k-th layer after processing by the Adaptive Receptive Field (ARF) module. This represents a deformable convolution with a learned offset. Represents the channel attention weight vector. This indicates element-wise multiplication. Let i represent the input feature map at the i-th scale. This indicates that it has a learning offset; The expression for the centrality score of the target detection result is: ； in, This represents the predicted centrality score at spatial location (x, y). This represents the function that takes the minimum value. This represents the function that takes the maximum value. This represents the lateral distance from the predicted point (x, y) to the left boundary of the predicted bounding box. This represents the lateral distance from the predicted point (x, y) to the right boundary of the predicted bounding box. This represents the vertical distance from the predicted point (x, y) to the upper boundary of the predicted bounding box. This represents the vertical distance from the predicted point (x, y) to the lower boundary of the predicted bounding box.

4. The multi-scale target detection method for airport scenes based on an enhanced target detection model according to claim 1, characterized in that, S3 includes: Using a multi-scale moving target dataset from an airport surface, small target enhancement was performed on the enhanced target detection model to obtain small target enhancement results; Based on the small target augmentation results, the size is adjusted using bilinear interpolation, and the sampling probability of small targets is dynamically increased using scale-aware mosaic sampling to obtain the augmented dataset; Based on the augmented dataset, a multi-target training method is used through a sample allocation mechanism to obtain a well-trained augmented target detection model.

5. The multi-scale target detection method for airport scenes based on an enhanced target detection model according to claim 4, characterized in that, The sample allocation mechanism includes: Based on the augmented dataset, a cost matrix is ​​constructed using the optimal transmission allocation strategy; By solving the cost matrix, the optimal matching result is obtained. The enhanced target detection model is then trained on multiple targets to obtain a well-trained enhanced target detection model.

6. The multi-scale target detection method for airport scenes based on an enhanced target detection model according to claim 5, characterized in that, The expression for the cost matrix is: ； in, This represents the value in the i-th row and j-th column of the cost matrix. The weights represent the classification loss. Indicates focal loss. This represents the probability distribution of the class predictions for the i-th prediction box. This represents the true category label of the j-th true bounding box. The weights representing the regression loss. Indicates CIoU loss. This represents the bounding box coordinates of the i-th predicted box by the model. Represents the bounding box coordinates of the j-th ground truth bounding box; The expression for the optimal matching result is: ； in, This represents the optimal matching result. This indicates that the constraints are met. All transmission plans P Among them, find the one that minimizes the objective function. P , Indicates the transportation plan, Representing the cost matrix, Representation matrix P and C The inner product, This represents the entropy regularization term.

7. The multi-scale target detection method for airport scenes based on an enhanced target detection model according to claim 4, characterized in that, The small target enhancement process includes: Extracting the area of ​​target objects from a multi-scale moving target dataset at airport surfaces; For each target object area, generate multiple duplicate samples; Apply an affine transformation to the duplicate samples and paste them to obtain the corresponding new coordinates; Based on the new coordinates of each repeated sample, image fusion is performed on the repeated samples to obtain the small target enhancement result.

8. The multi-scale target detection method for airport scenes based on an enhanced target detection model according to claim 7, characterized in that, The expression for the new coordinates is: ； in, Indicates the new center point coordinates. Represents the original x-coordinate. This represents the horizontal coordinate offset randomly generated within the image range. Represents the original position's ordinate. This represents the offset of the ordinate position randomly generated within the image range; The expression for the small target enhancement result is: ； in, This indicates that smaller goals enhance the results. Indicates the transparency blending factor. Represents the foreground image. This represents the background image.

9. The multi-scale target detection method for airport scenes based on an enhanced target detection model according to claim 4, characterized in that, The expression for the augmented dataset is: ； ； ； ； ； in, This represents the input scale randomly selected in the t-th training iteration. This indicates rounding to the nearest integer. This indicates the base resolution, which defaults to 640 pixels. This represents the scale factor, set to 0.

5. This represents a random number sampled from a uniform distribution U(0,1). Indicates uniform distribution. This represents the pixel value at position (x', y') of the resized image. and This represents the coordinates in the image after resizing. Indicates the position of the original image ( , The pixel value at () This represents the function that takes the maximum value. This indicates the adjusted image width. This indicates the adjusted image height. Indicates the scale of dynamic selection. Indicates the height of the original image. Indicates the width of the original image. This represents the spatial attention weight at position (i,j). This represents the sigmoid activation function. This represents the total number of channels in the feature map, and c represents the loop variable used to iterate through the channel indices. This represents the value of the feature map at position (i,j) and channel c. This indicates the probability that an image containing small objects was sampled. This indicates the number of images in the dataset that contain small objects. This represents the total number of images in the dataset.

Images