An unmanned aerial vehicle aerial image vehicle detection method based on improved YOLOv11

Abstract

The application discloses a kind of unmanned aerial vehicle aerial image vehicle detection methods based on improved YOLOv11, obtain unmanned aerial vehicle aerial image data, part or all aerial image contains vehicle;Utilize the image data training detection model, the detection model is based on YOLOv11 network, its main network extracts multiple scale feature maps, its neck network introduces semantic and detail injection module instead of original fusion unit;The semantic and detail injection module will be input feature map after it through spatial attention and channel attention processing, again after uniform size and smoothing processing, by hierarchical feature interaction explicit injection semantic information to bottom feature, while injecting high-level feature with fine-grained spatial details;Vehicle detection is carried out using the trained detection model.The application improves through multi-module collaborative, significantly improves the detection accuracy and robustness of small target and occluded target, while maintaining relatively low computational complexity and parameter amount.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and computer vision technology, and specifically relates to a vehicle detection method based on improved YOLOv11 for drone aerial images, which is applicable to smart cities, traffic monitoring, autonomous driving assistance and other scenarios. Background Technology

[0002] With the acceleration of urbanization and the popularization of new energy vehicles, the number of vehicles in cities has surged, and traffic networks have become increasingly complex, posing serious challenges to traffic control and management. As a new type of inspection tool, drones have the advantages of high mobility and flexible deployment, demonstrating great application potential in areas such as traffic flow monitoring, accident handling, and violation detection.

[0003] Object detection is a core technology in drone applications. In recent years, deep learning-based object detection methods, especially the YOLO series algorithms, have attracted much attention due to their good balance between detection accuracy and speed. However, drone aerial images present unique challenges: 1) Variable target scale: Changes in drone flight altitude lead to significant differences in the scale of the same target in the image; 2) Dense small targets: Distant targets such as pedestrians and bicycles occupy only a few pixels in the image, resulting in weak feature information; 3) Severe occlusion: Mutual occlusion between vehicles and pedestrians, as well as occlusion by buildings and trees, leads to incomplete target features. Existing YOLO models often suffer from missed detections and false detections when handling these complex scenes, making it difficult to meet the requirements of high-precision applications.

[0004] Therefore, how to improve the YOLO model's ability to detect small and occluded targets in UAV aerial images while maintaining its high efficiency is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention aims to provide a vehicle detection method for UAV aerial images based on an improved YOLOv11, which solves the problems of variable target scale, low detection accuracy of small targets and occluded targets in the prior art. Through multi-module collaborative improvement, the detection accuracy and robustness of small targets and occluded targets in aerial images are significantly improved, while maintaining low computational complexity and number of parameters.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A vehicle detection method based on improved YOLOv11 for drone aerial images includes the following steps: Step 1: Acquire drone aerial image data, in which some or all of the aerial images contain vehicles; Step 2: Train a detection model using the image data. The detection model is based on the YOLOv11 network. Its backbone network extracts feature maps at multiple scales, and its neck network introduces a semantic and detail injection module to replace the original fusion unit. The semantic and detail injection module processes its input feature map through spatial attention and channel attention, and then through size unification and smoothing. It then explicitly injects semantic information into the lower-level features through hierarchical feature interaction, while injecting fine-grained spatial details into the higher-level features. Step 3: Use the trained detection model to detect vehicles.

[0007] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Improved small target detection capability: Through deep fusion of semantic and detail injection modules and global-local information extraction of context-guided feature fusion module, the feature representation of small targets is effectively enhanced, preventing them from being "submerged" in feature transmission, and significantly improving the detection accuracy and localization accuracy of small targets.

[0008] 2. Enhanced robustness against occluded targets: Through the adaptive focusing of the convolutional block attention module and the information completion of the context-guided feature fusion module, the model can more effectively utilize the discriminative features of the visible part of the target and combine them with global semantic information for inference, so that robust recognition can be achieved even if the target is partially occluded.

[0009] 3. Maintains model efficiency: While improving accuracy, this invention controls the number of parameters and computational complexity at a low level through lightweight module design (such as SDI, CBAM, CGFM) and dynamic detection head, achieving a balance between high detection accuracy and high inference speed, making it suitable for deployment on resource-constrained edge devices.

[0010] 4. Comprehensive performance improvement: Experiments on public datasets show that the method of this invention has significant improvements in key indicators such as accuracy, recall, and mAP50 compared to the benchmark YOLOv11 model, especially in small target categories such as pedestrians and bicycles. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the overall framework and process of the present invention.

[0012] Figure 2 This is a schematic diagram of the CBAM attention model of the present invention.

[0013] Figure 3 This is a schematic diagram of the dynamic monitoring head architecture of the present invention.

[0014] Figure 4This is a comparison of the target recognition performance of YOLO-BiCCD and YOLOv11 in poor lighting conditions in an embodiment of the present invention. The left image is YOLO-BiCCD, and the right image is YOLOv11.

[0015] Figure 5 This is a comparison of the target recognition performance of YOLO-BiCCD and YOLOv11 under strong light conditions in this embodiment of the invention. The left image is YOLO-BiCCD and the right image is YOLOv11.

[0016] Figure 6 This is a comparison of the recognition performance of YOLO-BiCCD and YOLOv11 in dense target scenes in this embodiment of the invention. The left image is YOLO-BiCCD, and the right image is YOLOv11.

[0017] Figure 7 This invention presents a comparison of the target recognition performance of YOLO-BiCCD and YOLOv11 in complex scenes. The left image shows YOLO-BiCCD, and the right image shows YOLOv11. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] like Figure 1 As shown, the present invention discloses a vehicle detection method for UAV aerial images based on an improved YOLOv11, the main steps of which are as follows: Step 1, data preprocessing.

[0020] Acquire drone aerial imagery data, in which some or all of the images should contain vehicles. Specifically, tools such as Labelme, FlyPix AI, and ArcGIS Drone2Ma can be used, which support remote sensing / aerial photography-specific formats, have built-in AI-assisted annotation, and perform data enhancement, contrast and detail improvement, and slicing to form the model input data.

[0021] Slicing is an important method in this invention. Through slicing, this invention can solve the problem of inaccurate small target recognition. The specific processing method is as follows: The high-resolution image is first sliced ​​into several smaller images, which are then fed into the detection model. Finally, the detection results are stitched back together to form the original image. This method, also known as slice-assisted super inference, can significantly improve the recall rate for small targets.

[0022] Step 2: Establish a detection model based on the YOLOv11 network and extract multi-scale features through its backbone network.

[0023] This step involves inputting the image processed in step 1 into the YOLOv11 backbone network. Further, several C3k2 and C2PSA attention mechanisms can be introduced into the backbone network to achieve a dual improvement in feature extraction efficiency and accuracy. The outputs of the C3k2 and C2PSA modules yield feature maps at multiple scales. The entire processing flow can be understood as a process of progressively transforming the input image into multi-scale feature maps.

[0024] Step 3: Feature fusion and enhancement are performed by reconstructing the neck network of the detection model.

[0025] This invention improves the YOLOv11 neck network by introducing a Semantic and Details Injection (SDI) module to replace the original simple fusion unit. The SDI module processes its input feature map through spatial attention and channel attention, and then through size unification and smoothing. It then explicitly injects semantic information into the lower-level features through hierarchical feature interaction, while injecting fine-grained spatial details into the higher-level features.

[0026] In this invention, SDI, as an important feature fusion module, needs to receive different features and fuse them to improve the model's feature extraction capability and obtain richer feature information. Specifically, for M feature maps of different sizes... (Right now Figure 1 The feature fusion process of the featureMap0-M (in the semantic and detail injection modules) is as follows: The first step, as shown in formula (1), is to process the input feature map received by SDI. ( After processing with spatial attention and channel attention, the result is obtained. ,as follows: (1) Indicates the first i Layer channel attention processing, Indicates the first i Spatial attention processing of layers.

[0027] The second step is to use a 1×1 convolution to... The number of channels was uniformly reduced to c ,get The size is .

[0028] The third step requires adding the first [item] to the decoder. i The layer generates a fused feature, assuming the feature map of this layer itself has a resolution of . For any other layer, based on any feature map obtained in the second step... (Its resolution may be larger, smaller, or the same), to unify its size to Size get The method is as shown in formula (2): (2) in, For adaptive average pooling; For identity mapping, the feature map remains unchanged without any processing; This is a bilinear interpolation.

[0029] The fourth step, as shown in formula (3), is to use a 3×3 convolution to process the resized feature map. Smooth the surface to obtain the desired result. .

[0030] (3) in, For smooth convolution.

[0031] The fifth step involves explicitly injecting semantic information into lower-level features through hierarchical feature interaction, while simultaneously injecting fine-grained spatial details into higher-level features to obtain fused features. This ensures that each feature layer possesses both rich semantic understanding and precise detail representation. Its core operation can be formalized as formula (4): (4) in, This represents the Hadamard product of the principal elements of the feature maps across different levels. This is a smooth convolution operator.

[0032] Furthermore, in Figure 1 In the illustrated embodiment, the backbone network outputs feature maps at four scales: lowest-level feature map, second-lowest-level feature map, and highest-level feature map. The number of semantic and detail injection modules is two in this embodiment. In this embodiment, the detection model of the present invention also includes three context-guided feature fusion modules (CGFM), where: The lowest-level feature map, after convolution processing, is input into the first semantic and detail injection module; The second-lowest level feature map is divided into two paths after convolution processing. One path is input to the first semantic and detail injection module, and the other path is input to the first CGFM module. The sub-high-level feature maps are divided into two paths after convolution processing. One path is input to the second semantic and detail injection module, and the other path is input to the second CGFM module. The highest-level feature map is divided into two paths after being processed by the CBAM attention module and convolution. One path is upsampled and then input into the second CGFM module to be fused with the second-highest-level feature map, while the other path is input into the third CGFM module. The fusion features of the second CGFM module are processed by the C3K2 module and then divided into two paths. One path is upsampled and input into the first CGFM module to fuse with the second low-level feature map, and the other path is input into the second semantic and detail injection module. The fused features of the first CGFM module are processed by the C3K2 module and then input separately into the first semantic and detail injection module, where they are fused with the second-lowest level feature map and the lowest level feature map. The fusion features of the first semantic and detail injection module are divided into two paths after being processed by the C3K2 module and convolution. One path is used as the first input feature of the head network, and the other path is input to the second semantic and detail injection module, where they are fused with the fusion features of the second-highest level feature map and the second CGFM module. The fusion features of the second semantic and detail injection module are processed by the C3K2 module and then divided into two paths. One path serves as the second input feature of the head network, and the other path is processed by convolution and then input into the third CGFM module to be fused with the highest level feature map. The fusion features of the third CGFM module are processed by the C3K2 module and used as the third input features of the head network.

[0033] The CGFM module of this invention uses a dynamic gating mechanism to control its two input feature maps (high-level features). and low-level features The feature fusion process involves complementary weighted fusion, followed by recalibration of the fused features using an efficient channel attention mechanism. The main steps of feature fusion are as follows: The first step is to and Align them to the same channel dimension to obtain and .

[0034] Since the two input feature maps may have different numbers of channels, they need to be aligned to the same dimension first for subsequent fusion operations. Formula (5) is used to align the channels respectively. High-level characteristics (Rich in semantic information) and the number of channels is Low-level features (Rich in spatial detail) Alignment ensures that the channel dimensions of the two images match, avoiding dimensional conflicts during stitching or weighting. This process is represented as follows: Define the alignment function: (5) in The number of channels represents the feature map, and the target number of channels is... , It is 1 1. Convolution, used to adjust the number of channels. for The number of channels, set equal Number of channels Output , .

[0035] The second step involves generating dynamic weights G through a gating mechanism to determine... and The fusion ratio.

[0036] As shown in formula (6), a dynamic weight G is generated through a gating mechanism to determine the fusion ratio of the two feature maps. The weight G is a two-dimensional weight map with the same spatial size as the feature map, and its value ranges from (0,1). Its function is to determine whether each spatial location should rely more on the contribution of semantic features or the contribution of spatial details. Its calculation formula is: (6) in yes Sigmoid Activation function Concat This indicates the splicing of channel dimensions.

[0037] The third step is to use dynamic weights G to... and Perform weighting to obtain the weighted features. and This achieves dynamic fusion, ensuring that the information in the two feature maps is complementary rather than redundant. The process can be expressed as formulas (7) and (8), where... This represents element-wise multiplication (Hadamard product).

[0038] (7) (8) The fourth step involves concatenating the weighted features and introducing an attention enhancement mechanism for channel attention recalibration, followed by introducing residual connections. and Then, the final output is obtained by splicing. .

[0039] Specifically, the weighted features are first concatenated using formula (9). (9) Then, as in formula (10), the attention enhancement mechanism (ECA Attention formula (11)) is used to... Enhanced features are obtained by channel attention recalibration. This will further enhance the response of critical channels.

[0040] (10) The attention enhancement mechanism here strengthens the response of key channels by introducing an efficient channel attention mechanism that adaptively recalibrates channel weights. Perform global average pooling, then perform 1D convolution (kernel size k) to capture local cross-channel interactions, generate channel attention weights, and apply them to the fused feature map, as follows: (11) in This represents a 1×1 convolution with a kernel size of k.

[0041] Finally, residual connections are used to preserve the original features, avoiding information loss and ensuring smooth gradients. Then, the enhanced features are concatenated to obtain the final output. This process can be expressed as formula (12): (12) Furthermore, this invention introduces a CBAM attention mechanism at key fusion points, generating adaptive weights in both channel and spatial dimensions to guide the network to focus its attention on the most informative feature channels and spatial regions. Addressing the pain points of easily lost features for small targets and incomplete features for occluded targets, the CGFM feature fusion module is improved. Through refined contextual guidance, it deeply fuses multi-scale features, extracting complementary information from the global and local environments, allowing high-level semantics and low-level details to mutually enhance each other. This pixel-level feature interaction effectively completes the information of small targets and occluded targets, significantly improving their detectability.

[0042] Specifically, the CBAM attention module includes a cascaded channel attention module and a spatial attention module; the channel attention module aggregates spatial information through global average pooling and global max pooling, and generates channel weights via a shared multilayer perceptron; the spatial attention module performs average pooling and max pooling along the channel dimension, and generates spatial weights via convolutional layers, such as... Figure 2 As shown.

[0043] The channel attention formula is: The spatial attention formula is: in and This represents the features after average pooling. and Represents the features after max pooling. , The weights represent the multilayer perceptron (MLP) weights. and The inputs are shared, and the ReLU activation function is followed by... ; W represents the Sigmoid function. 0 ∈R C×C / r MLP stands for Multilayer Perceptron. This represents a convolution operation with a 7×7 convolution kernel.

[0044] Step 4, detection.

[0045] The detection model of this invention improves the head network by using the Dyhead dynamic detection head with an integrated dynamic attention mechanism. The dynamic detection head dynamically adjusts the receptive field and feature response through scale-aware, spatial-aware, and task-aware attention mechanisms to generate the final detection result, including the target category and bounding box coordinates.

[0046] Specifically, the dynamic detection head of this invention integrates three dynamic attention mechanisms for dynamically aggregating information from different feature pyramid levels, enabling adaptive processing of different spatial locations, scales, and task requirements in the feature map, which are described below: 1) Scale-aware attention dynamically aggregates information from different feature pyramid levels, as shown in the following expression: in f ( ) is a linear function approximated by a 1×1 convolutional layer. It is the hard-sigmoid function. H and W represent the size of the feature map, C represents the number of channels in the feature map, and L represents the number of different scales of the input features, i.e., the number of layers in the feature pyramid.

[0047] 2) Spatial awareness attention employs dynamically deformable convolution to dynamically predict sampling offsets based on local content of input features, adaptively adapting to terrain changes to accurately capture target contours. The expression is as follows: Where K represents the number of coefficient sampling points. Through self-learning space offset The adjusted position coordinates are used to focus on the discriminative region; It is a location The importance scalar for self-learning, and both are learned from the input features of the intermediate feature F; L represents the number of feature layers. Representing the l The weights of sampling points in layer k.

[0048] 3) Task-aware attention employs a dynamic activation function to provide adaptive nonlinear transformations for classification and regression tasks, as shown in the following expression: in Indicates the first c Feature slices of each channel It is a superfunction that controls the activation threshold through learning.

[0049] Through these mechanisms, the detection head can perform feature sampling and prediction in a more flexible and intelligent way, thereby achieving more accurate localization and classification, with a structure such as... Figure 3 As shown.

[0050] Step 5, Model Training.

[0051] This invention optimizes the loss function during training, specifically by employing the MPDIoU loss function for bounding box regression. The MPDIoU loss function optimizes the regression accuracy of the bounding boxes by minimizing the distances between the top-left and bottom-right corners of the predicted and ground truth bounding boxes. Specifically: The MPDIoU loss function is expressed as follows: The MPDIoU similarity metric is expressed as follows: It is the intersection-union ratio (IoU) between the predicted bounding boxes and the ground truth bounding boxes, expressed as follows: It is the square of the Euclidean distance between the top-left corner of the predicted bounding box and the top-left corner of the ground truth bounding box, expressed as follows: It is the square of the Euclidean distance between the bottom right corner of the predicted bounding box and the bottom right corner of the ground truth bounding box, expressed as follows: in Represents the true frame. Represents the prediction box. This represents the Euclidean distance between the top-left corner coordinates of the ground truth bounding box and the predicted bounding box. This represents the Euclidean distance between the bottom right corner coordinates of the ground truth bounding box and the predicted bounding box. , This represents the coordinates of the top-left corner of the true bounding box. , This represents the coordinates of the bottom right corner of the actual bounding box. , This represents the coordinates of the top-left corner of the prediction box. , The coordinates of the bottom right corner of the prediction box are represented by h and w, which represent the height and width of the input graph.

[0052] In one specific embodiment of the present invention, the hardware environment is a server configured with an NVIDIA GeForce RTX4090 GPU, and the software environment is an Ubuntu 22.04 operating system, Python 3.10, and the PyTorch deep learning framework.

[0053] This embodiment uses the publicly available VisDrone2019 drone aerial photography dataset for model training and testing. This dataset contains a large number of images under different scenes and lighting conditions, with targets including pedestrians, cars, bicycles, tricycles, etc., and contains a large number of small targets and occluded targets, making it very suitable for verifying the effectiveness of the present invention.

[0054] The model training process is as follows: Images in the VisDrone2019 dataset were uniformly scaled to 640×640 pixels and subjected to data augmentation techniques such as random flipping and color jittering. A model architecture based on an improved YOLOv11 was used, and the network was constructed following the steps outlined in the above technical solution. Training was performed using the SGD optimizer with an initial learning rate of 0.01, momentum of 0.937, and batch size of 32, for a total of 300 epochs. The MPDIoU loss function was used for bounding box regression, and the cross-entropy loss function was used for classification.

[0055] Figure 4 , Figure 5 , Figure 6 and Figure 7 The presentation compares some detection results of the proposed model (YOLO-BiCCD) with the baseline model (YOLOv11) on the VisDrone 2019 test set. It is evident that in complex scenes such as low light, strong light, dense lighting, and occlusion, the proposed model can detect more small targets, and its identification of occluded targets is more accurate, with higher detection confidence.

[0056] Table 1 shows the performance comparison of the model of this invention with current mainstream YOLO series models and other advanced models on the VisDrone2019 dataset. It can be seen that the model of this invention achieves the highest overall mAP of 49.44%, significantly outperforming models such as YOLOv11 (44.20%) and YOLOv8l (46.82%). Simultaneously, it maintains excellent performance in terms of parameter count (2.66M) and inference speed (493.41 FPS), verifying the advanced nature and practicality of this invention.

[0057] Table 1. Performance Comparison of Other YOLO Series and Improved Models on Different Objectives YOLOv4 24.80% 12.60% 8.60% 64.30% 22.40% 22.70% 11.40% 7.60% 44.30% 21.70% 24.04% YOLOv5s 39.80% 31.70% 10.20% 73.00% 32.60% 27.10% 19.30% 9.70% 40.50% 38.10% 32.20% YOLOv8l 51.10% 39.80% 21.90% 82.90% 49.30% 45.40% 38.10% 20.40% 67.00% 52.30% 46.82% YOLOv11 43.57% 51.01% 21.97% 64.77% 45.03% 45.16% 38.97% 29.32% 58.97% 43.21% 44.20% YOLOv5l 47.80% 37.70% 17.80% 78.20% 42.60% 40.30% 26.80% 13.10% 54.70% 45.30% 40.43% PDWT-YOLO 48.70% 41.60% 14.70% 82.00% 43.20% 35.40% 26.80% 14.20% 56.40% 49.30% 41.23% DDSC-YOLO 48.90% 40.30% 15.10% 83.10% 47.80% 35.00% 27.60% 16.30% 59.10% 48.60% 42.18% LUMF-YOLO 51.90% 40.50% 21.40% 80.50% 46.30% 40.90% 30.20% 15.70% 57.10% 48.60% 43.31% Tph-YOLOv5 29.00% 16.75% 15.69% 68.94% 49.79% 45.16% 27.33% 24.72% 61.80% 30.90% 37.01% YOLOv3-SPP 49.40% 39.40% 17.80% 78.10% 42.80% 37.90% 26.70% 14.30% 50.60% 45.60% 40.26% YOLO-BiCCD 50.16% 57.03% 29.11% 69.26% 48.14% 51.48% 44.76% 32.32% 62.36% 49.80% 49.44% In summary, this invention proposes a vehicle detection method for UAV aerial images based on an improved YOLOv11 model. The main improvements are to the neck and head networks of the YOLOv11 model. The neck network is improved by replacing the original fusion units with semantic and detail injection modules, achieving efficient interaction and deep fusion of cross-level features. Convolutional block attention modules are introduced at key fusion points to adaptively enhance feature discrimination from both channel and spatial dimensions. Furthermore, a context-guided feature fusion module is introduced to extract global-local complementary information, achieving bidirectional enhancement of high- and low-level features. The head network is improved by integrating a dynamic detection head, dynamically optimizing feature responses by combining scale, spatial, and task-aware mechanisms. By improving the feature fusion network and dynamic detection head, this invention effectively solves the problems of difficult detection of small targets and occluded targets in aerial images, maintaining model efficiency while improving detection accuracy, and providing strong technical support for fields such as intelligent transportation and security monitoring.

[0058] While the present invention has been described through preferred embodiments, it is not limited to the embodiments described herein, and various changes and modifications are made without departing from the scope of the invention.

Claims

1. A vehicle detection method based on improved YOLOv11 in UAV aerial images, characterized in that, Includes the following steps: Step 1: Acquire drone aerial image data, in which some or all of the aerial images contain vehicles; Step 2: Train a detection model using the image data. The detection model is based on the YOLOv11 network. Its backbone network extracts feature maps at multiple scales, and its neck network introduces a semantic and detail injection module to replace the original fusion unit. The semantic and detail injection module processes its input feature map through spatial attention and channel attention, and then through size unification and smoothing. It then explicitly injects semantic information into the lower-level features through hierarchical feature interaction, while injecting fine-grained spatial details into the higher-level features. Step 3: Use the trained detection model to detect vehicles.

2. The vehicle detection method for UAV aerial images based on improved YOLOv11 according to claim 1, characterized in that, Step 1 involves preprocessing the acquired aerial images, including slicing, which is as follows: First, the high-resolution large image is cut into several smaller images, which are then fed into the detection model. Finally, the detection results are stitched back together to form the original image.

3. The vehicle detection method for UAV aerial images based on improved YOLOv11 according to claim 1, characterized in that, For M feature maps of different sizes The feature fusion process of the semantic and detail injection module is as follows: The first step is to input the feature map. After processing with spatial attention and channel attention, the result is obtained. ,as follows: Indicates the first i Layer channel attention processing, Indicates the first i Spatial attention processing of layers; The second step is to use a 1×1 convolution to... The number of channels was uniformly reduced to c ,get ; The third step is to use the arbitrary feature map obtained in the second step. Size uniformity Size get The method is as follows: in, For adaptive average pooling; For identity mapping, the feature map remains unchanged without any processing; It is a bilinear interpolation; Fourth step, use 3×3 convolution to process the resized feature map. Smooth the surface to obtain the desired result. The formula is as follows: in, For smooth convolution; The fifth step involves explicitly injecting semantic information into lower-level features through hierarchical feature interaction, while simultaneously injecting fine-grained spatial details into higher-level features to obtain fused features. As shown in the following formula: in, This represents the Hadamard product of the principal elements across the feature maps.

4. The vehicle detection method for UAV aerial images based on improved YOLOv11 according to claim 1 or 3, characterized in that, The backbone network outputs feature maps at four scales. There are two semantic and detail injection modules. The detection model also includes three CGFM modules, wherein: The lowest-level feature map, after convolution processing, is input into the first semantic and detail injection module; The second-lowest level feature map is divided into two paths after convolution processing. One path is input to the first semantic and detail injection module, and the other path is input to the first CGFM module. The sub-high-level feature maps are divided into two paths after convolution processing. One path is input to the second semantic and detail injection module, and the other path is input to the second CGFM module. The highest-level feature map is divided into two paths after being processed by the CBAM attention module and convolution. One path is upsampled and then input into the second CGFM module to be fused with the second-highest-level feature map, while the other path is input into the third CGFM module. The fusion features of the second CGFM module are processed by the C3K2 module and then divided into two paths. One path is upsampled and input into the first CGFM module to fuse with the second low-level feature map, and the other path is input into the second semantic and detail injection module. The fused features of the first CGFM module are processed by the C3K2 module and then input separately into the first semantic and detail injection module, where they are fused with the second-lowest level feature map and the lowest level feature map. The fusion features of the first semantic and detail injection module are divided into two paths after being processed by the C3K2 module and convolution. One path is used as the first input feature of the head network, and the other path is input to the second semantic and detail injection module, where they are fused with the fusion features of the second-highest level feature map and the second CGFM module. The fusion features of the second semantic and detail injection module are processed by the C3K2 module and then divided into two paths. One path serves as the second input feature of the head network, and the other path is processed by convolution and then input into the third CGFM module to be fused with the highest level feature map. The fusion features of the third CGFM module are processed by the C3K2 module and used as the third input features of the head network.

5. The vehicle detection method for UAV aerial images based on improved YOLOv11 according to claim 4, characterized in that, The two input features of the CGFM module are high-level features. and low-level features The feature fusion process is as follows: The first step is to and Align them to the same channel dimension to obtain and ; The second step involves generating dynamic weights through a gating mechanism to determine... and The fusion ratio; The third step is to utilize the aforementioned dynamic weights. and We perform weighting to obtain the weighted features. and ; The fourth step involves concatenating the weighted features and introducing an attention enhancement mechanism for channel attention recalibration, followed by introducing residual connections. and Then, the final output is obtained by splicing. .

6. The vehicle detection method for UAV aerial images based on improved YOLOv11 according to claim 5, characterized in that, The first step is to define the alignment function. in The number of channels represents the feature map, and the target number of channels is... , It is 1 1. Convolution, used to adjust the number of channels. for The number of channels, set equal Number of channels Output , ; The second step is dynamic weighting. The calculation formula is: in yes Sigmoid Activation function Concat This indicates the concatenation of channel dimensions; The third step, the weighting formula, is as follows: The fourth step involves first concatenating the weighted features using the following formula: Then, attention enhancement mechanisms were used to... Enhanced features are obtained by channel attention recalibration. ; Finally, the original features are preserved through residual connections, and the enhanced features are concatenated to obtain the final output, as shown in the following equation: 。 7. The vehicle detection method for UAV aerial images based on improved YOLOv11 according to claim 4, characterized in that, The CBAM attention module includes a cascaded channel attention module and a spatial attention module; the channel attention module aggregates spatial information through global average pooling and global max pooling, and generates channel weights through a shared multilayer perceptron; the spatial attention module performs average pooling and max pooling along the channel dimension, and generates spatial weights through a convolutional layer. The channel attention formula is: The spatial attention formula is: in and This represents the features after average pooling. and Represents the features after max pooling. , The weights represent the multilayer perceptron (MLP) weights. and The inputs are shared, and the ReLU activation function is followed by... ; W represents the Sigmoid function. 0 ∈R C×C / r MLP stands for Multilayer Perceptron. This represents a convolution operation with a 7×7 convolution kernel.

8. The vehicle detection method for UAV aerial images based on improved YOLOv11 according to claim 1, characterized in that, The head network employs the Dyhead dynamic detection head, which integrates a dynamic attention mechanism. This dynamic detection head dynamically adjusts the receptive field and feature response through scale-aware, spatial-aware, and task-aware attention mechanisms to generate the final detection result, including the target category and bounding box coordinates.

9. The vehicle detection method for UAV aerial images based on improved YOLOv11 according to claim 8, characterized in that, The scale awareness dynamically aggregates information from different feature pyramid levels, as shown in the following expression: in f ( ) is a linear function approximated by a 1×1 convolutional layer. It is the hard-sigmoid function. H and W represent the size of the feature map, C represents the number of channels in the feature map, and L represents the number of different scales of the input features, i.e., the number of layers in the feature pyramid. The spatial awareness employs dynamically deformable convolution to dynamically predict sampling offsets based on local content of input features, adaptively adapting to terrain changes to accurately capture target contours. The expression is as follows: Where K represents the number of coefficient sampling points. Through self-learning space offset The adjusted position coordinates are used to focus on the discriminative region; It is a location The importance scalar for self-learning, and both are learned from the input features of the intermediate feature F; L Represents the number of feature layers. Representing the l The weights of the sampling points in layer k; The task awareness employs a dynamic activation function to provide adaptive nonlinear transformations for classification and regression tasks, as expressed below: in Indicates the first c Feature slices of each channel It is a superfunction that controls the activation threshold through learning.

10. The vehicle detection method for UAV aerial images based on improved YOLOv11 according to claim 1, characterized in that, The detection model uses the MPDIoU loss function for bounding box regression during training. The MPDIoU loss function optimizes the bounding box regression accuracy by minimizing the distance between the top-left corner and the bottom-right corner of the predicted box and the ground truth box. The MPDIoU loss function is expressed as follows: The MPDIoU similarity metric is expressed as follows: It is the intersection-union ratio (IoU) between the predicted bounding boxes and the ground truth bounding boxes, expressed as follows: It is the square of the Euclidean distance between the top-left corner of the predicted bounding box and the top-left corner of the ground truth bounding box, expressed as follows: It is the square of the Euclidean distance between the bottom right corner of the predicted bounding box and the bottom right corner of the ground truth bounding box, expressed as follows: in Represents the true frame. Represents the prediction box. This represents the Euclidean distance between the top-left corner coordinates of the ground truth bounding box and the predicted bounding box. This represents the Euclidean distance between the bottom right corner coordinates of the ground truth bounding box and the predicted bounding box. , This represents the coordinates of the top-left corner of the true bounding box. , This represents the coordinates of the bottom right corner of the actual bounding box. , This represents the coordinates of the top-left corner of the prediction box. , The coordinates of the bottom right corner of the prediction box are represented by h and w, which represent the height and width of the input graph.

Images